APPROVED: 1 April 2022 doi:10.2903/sp.efsa.2022.EN-7339 Roadmap for actions on artificial intelligence for evidence management in risk assessment Final Report Bersani C., Codagnone J., David L., Foiniotis A., Galasso G., Mancini S., Michieletti R., Orphanidou C., Pellegrino M. PwC EU Services & Intellera Consulting Srl Abstract In May 2021, the European Food Safety Authority (EFSA) launched the “Roadmap for actions on Artificial Intelligence for evidence management in risk assessment” project to develop an approach for the implementation of Artificial Intelligence (AI) methods in the evidence management phase of its internal risk assessment process. The main objective is to identify specific projects to be carried out to increase – by 2027 - the accessibility and breadth of the body of evidence and enhance the trustworthiness of the risk assessment process by applying human-centric AI in close coexistence with human expertise. The roadmap, therefore, presents a full understanding of ongoing activities, market readiness, knowledge gaps, and societal interests and concerns, as well as collaboration opportunities in the field of AI application to the evidence management process. Subsequently, based on these findings, it outlines a set of recommendations aimed at developing the Agency's capabilities to adopt and integrate AI solutions in the evidence management process. Finally, the roadmap contains considerations aimed at identifying potential communication and engagement opportunities in the area of AI applied to evidence management in risk assessment, considering specific needs and concerns of EFSA's stakeholders on the adoption of AI. © European Food Safety Authority, 2022 Key words: Evidence Management, Risk Assessment, Artificial Intelligence (AI), Natural Language Processing (NLP), Machine Learning (ML), EFSA Question number: EFSA-Q-2022-00230 Correspondence: SPIDO@efsa.europa.eu www.efsa.europa.eu/publications EFSA Supporting publication 2022: EN-7339 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License EXTERNAL SCIENTIFIC REPORT Disclaimer: This document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. This document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view, and position as regards the issues addressed and the conclusions reached in this document, without prejudice to the rights of the authors. Acknowledgements: We would like to pay our special regards to Konstantinos Paraskevopoulos, Ermanno Cavalli, Angelo Cafaro and the other members of the EFSA SPIDO team for their guidance and support along the entire project, and to all EFSA staff members who provided insights and feedback to help us raise the quality of our work. We are also grateful to the external stakeholders who provided us with information during the data collection activities. Amendment: This version was revised by adding information on the Terms of Reference of the contract, as requested by the EFSA template for external scientific reports (page 5). This editorial correction does not materially affect the contents or outcome of this scientific output. To avoid confusion, the original version of the output has been removed from the EFSA Journal but is available on request. Suggested citation: PwC EU Services & Intellera Consulting, 2022. Roadmap for action on Artificial Intelligence for evidence management in risk assessment. EFSA supporting publication 2022: EN-7339. 120 pp. doi:10.2903/sp.efsa.2022.EN-7339 ISSN: 2397-8325 © European Food Safety Authority, 2022 Reproduction is authorised provided the source is acknowledged. www.efsa.europa.eu/publications 2 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Executive Summary The European Food Safety Authority (EFSA) is the Agency of the European Union that provides independent scientific advice and communicates on existing and emerging risks associated with the food chain. Evidence management is at the core of the evidence-based risk assessment: it is a structured process of collecting and analysing data generated through different types of studies and approaches, which are gathered by various sources. Evidence management plays a crucial role as it contributes to the development of EFSA's scientific opinions. However, it is a resource-intensive activity, where the amount of information processed is inherently limited by the accessibility to the body of evidence and to available resources. Given the relevance of the evidence management process within the organisation and the specific challenges faced in terms of efficiency, quality and transparency, EFSA has recently presented the Theme (concept) paper on Artificial Intelligence in risk assessment (2021) 1 with its strategic vision to build a harmonised approach on the implementation of artificial intelligence (AI) methods in the evidence management phase of the risk assessment process. In this theme paper, EFSA stated the objectives to be achieved by 2027: increasing the accessibility and breadth of the body of evidence and improving the trustworthiness in the risk assessment process, while applying human-centric AI in close coexistence with human expertise. In order to reach these objectives, EFSA launched the project Roadmap for actions on artificial intelligence for evidence management in risk assessment (specific contract No 35 – Rc 5 implementing Framework Contract No OC/EFSA/FIN/2019/01), awarded to PwC EU Services EESV and the subcontractor Intellera Consulting Srl. The purpose of the project is to deliver a roadmap for action, with recommendations for future multi-annual and multipartner studies or projects in the area of AI. This roadmap builds on EFSA´s vision and supports its preparedness for future risk assessments, while simultaneously preventing possible discrepancies on sensitive matters. In order to guide the activities for the roadmap development, the evidence management process was analysed and split up into subsequent phases, which were defined as use cases on which to focus the following activities. Then, to develop a complete and accurate roadmap, an analysis of the current market for AI solutions was conducted through various data collection activities – including desk research, literature review, a survey, and several semistructured interviews. These preliminary findings highlighted the potential for applying AI to specific evidence management use cases, given the availability of a relevant number of AI-based solutions in the market. Even though the technology readiness level of these solutions varies, the majority of them are currently in a commercialised stage, while others should be further developed and adapted to the agency's specific needs, and would also require an upgrade of the agency's data capabilities. Most of the identified solutions were developed by small/medium software companies, with a few solutions by academic organisations, such as universities and research centres. The AI-based solutions were then categorised according to the previously identified use cases of the evidence management. The categorisation made clear that the highest number of available AI tools can be integrated for the automation of the primary source for data collection of evidence management, which is the literature review. In order to explore the specific challenges faced by organisations similar to EFSA and draw upon best-practices with regard to AI development, the analysis also focused on the strategies, roadmaps and projects launched by institutional organisations. The data showed that only a few public organisations have applied AI within their evidence management process. Nevertheless, some relevant institutions have demonstrated their interest in investing in this area, leaving EFSA some ground for future collaborations. In general, the analysis showed that several opportunities to collaborate with different stakeholders might be available in order to develop AI-related projects. After obtaining a clear picture of the current state of the art, the next step was to identify the potential challenges and blockers that public organisations face when applying AI-based solutions. What emerged from this analysis is that the organisation should adopt a holistic organisational transformation, including for example the re-design of specific process and workflows, to successfully adopt AI. In addition, a gap analysis conducted on EFSA’s data 1 European Food Safety Authority, Theme (concept) paper – on Artificial intelligence in risk assessment (2020) (unpublished information) www.efsa.europa.eu/publications 3 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence maturity assessment highlighted potential challenges related to both EFSA’s internal capabilities and technical requirements. The aforementioned activities led to the identification of a comprehensive set of potential collaboration opportunities during the AI lifecycle – which involves various stages of the AI development and integration – with different stakeholder categories from the industrial, academic and institutional arenas. Finally, in order to specifically advise EFSA on the implementation of AI solutions in the evidence management process, a set of horizontal and vertical recommendations has been provided. On one hand, “horizontal recommendations” were developed to guide EFSA in building the technical, procedural and governance capabilities to improve its preparedness towards the adoption of AI. These were focused on the need to develop a sound data governance and infrastructure framework, together with Development and Operations Management (DevOps), in order to increase the agency's ability to manage, control and deliver AI applications and services at a high speed. In addition, one of the suggested recommendations is to adopt a Trustworthy AI framework – including the development of a Code of Conduct and a Risk Framework - to ensure that EFSA is able to address the ethical, social and legal challenges surrounding AI, in light of the potential impact of the AI Regulation proposed by the Commission. Furthermore, domain-specific ontologies need to be developed to support the integration of precise AI solutions for the automation and improvement of the evidence management process. Finally, the acquisition of specific technical competences for the development and maintenance of AI-based systems is highly recommended. On the other hand, “vertical recommendations” were provided to link each of the prioritised use cases within EFSA’s evidence management process with the optimal approach for procuring or developing AI solutions with key stakeholders. Furthermore, to specifically support EFSA in its decision-making and resource allocation processes, these recommendations (or projects) have been first evaluated in terms of budget and duration, and then prioritised taking into account their urgency and inter-dependencies. Based on this activity, projects have been distributed on a timeline, from 2022 to 2027. Finally, a communication and engagement plan has been developed to inform the Agency's stakeholders and its surrounding ecosystem, strengthen current partnerships, and foster new collaboration opportunities. Overall, this roadmap serves as a decision-making tool to identify and prioritise AI projects aimed to optimise EFSA’s evidence management process through the introduction and use of AI. By implementing the recommendations/projects devised in this roadmap, EFSA will be able to develop the fundamental competences that enable the effective integration of AI within the organisation, both in terms of technical requirements to manage, control, and use AI responsibly, and in terms of new skills and cultural change. These projects are particularly relevant for building the Agency’s capabilities to integrate AI tools within evidence management and other processes, thus promoting synergies with other roadmaps (e.g., the EU Partnership for next generation, systems-based Environmental Risk Assessment - PERA - and the New Approach Methodologies - NAMs). www.efsa.europa.eu/publications 4 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Table of Contents Abstract ........................................................................................................................................ 1 Executive Summary ....................................................................................................................... 3 International Consumer Protection and Enforcement Network .......................................................... 9 1. Introduction ..................................................................................................................... 11 2. Background ...................................................................................................................... 11 2.1. Policy Context .................................................................................................................. 12 2.2. EFSA's specific objectives .................................................................................................. 13 2.3. High level project phases .................................................................................................. 14 2.4. Problem formulation ......................................................................................................... 15 3. State of the art and analysis.............................................................................................. 20 3.1. Market overview, map of relevant activities and further areas ............................................. 21 3.2. Challenges and blockers related to the adoption of AI in evidence management ................... 29 3.3. Cooperation/collaboration opportunities ............................................................................. 34 4. Recommendations for projects .......................................................................................... 49 4.1. Horizontal Recommendations ............................................................................................ 50 4.2. Vertical Recommendations ................................................................................................ 75 4.3. Prioritisation of Working Areas .......................................................................................... 89 5. Communication and Engagement Plans.............................................................................. 96 5.1. General Approach ............................................................................................................. 96 5.2. Inform & Monitor/Listen Matrix Quadrant ......................................................................... 100 5.3. Communication Plan ....................................................................................................... 101 5.4. Engagement Plan ........................................................................................................... 107 5.5. Final Considerations on Communication and Engagement – Timeline and dependencies ..... 117 6. Conclusions .................................................................................................................... 118 www.efsa.europa.eu/publications 5 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence List of Figures Figure 1 – EU’s approach to AI ........................................................................................................................ 12 Figure 2 - Main activities of the project ............................................................................................................ 15 Figure 3 - Evidence management process phases ............................................................................................. 16 Figure 4 – Evidence management process phases and use cases ....................................................................... 18 Figure 5 – MCDA criteria for use case prioritisation ........................................................................................... 19 Figure 6 - Revised prioritisation of the use cases............................................................................................... 20 Figure 7 - Data collection activities ................................................................................................................... 23 Figure 8 - Use case relevance .......................................................................................................................... 23 Figure 9 - Number of solutions per use case and readiness level ........................................................................ 24 Figure 10 - Summary of key insights ................................................................................................................ 24 Figure 11 - General Challenges ........................................................................................................................ 30 Figure 12 - Gap Analysis of the EFSA ecosystem ............................................................................................... 31 Figure 13 - AI lifecycle phases ......................................................................................................................... 34 Figure 14 - State of Play of EFSA AI ecosystem of stakeholders ......................................................................... 36 Figure 15 - AI lifecycle and collaboration models .............................................................................................. 39 Figure 16 - Vendor selection and procurement management process ................................................................. 42 Figure 17 - Prioritisation matrix of institutional stakeholders ............................................................................. 45 Figure 18 - Vertical and Horizontal recommendations link .................................................................................. 49 Figure 19 - Link of Recommendations with AI Lifecycle Phases .......................................................................... 50 Figure 20 - Dimensions and horizontal recommendations .................................................................................. 50 Figure 21 - Indicative Data Infrastructure for EFSA ........................................................................................... 55 Figure 22 - Indicative Data Infrastructure for EFSA based on the Azure cloud platform. ...................................... 56 Figure 23- Questions that a DG framework answers .......................................................................................... 59 Figure 24 - Indicative DG Framework incorporating Tools & Technology, Process & Operating Model and People & Culture ........................................................................................................................................................... 59 Figure 25 - DG roadmap with increasing levels of maturity ................................................................................ 60 Figure 26 - DevOps Benefits ............................................................................................................................ 62 Figure 27 - ALTAI example of questions which refer to the human agency and oversight principles ..................... 69 Figure 28 - ETAPAS Indicator framework design and structure........................................................................... 70 Figure 29 - Steps for Data Science Competences Development .......................................................................... 73 Figure 30 - Six principles of People-centred Change and Adoption ..................................................................... 73 Figure 31 - Six pillars of change management approach .................................................................................... 74 Figure 32 - General Approach for Communication and Engagement ................................................................... 97 Figure 33 Communication/engagement strategies for the identified stakeholder segments (Impact/Relevance/Easiness to collaborate matrix) ........................................................................................... 99 Figure 34 - Communication/Engagement strategies for the identified stakeholders segments ............................ 100 Figure 35 - Inform & Monitor/Listen Quadrant ................................................................................................ 101 Figure 36 - Ad-hoc Communication Quadrant ................................................................................................. 101 Figure 37 - Maintain & Reinforce Collaboration and Engage in the ecosystem ................................................... 108 List of Tables Table Table Table Table Table Table 1 - Risk Assessment Workflow and Evidence Management Process ........................................................... 16 2 - Top 10 prioritised use-cases .............................................................................................................. 20 3 - Extensive list of initiatives ................................................................................................................. 21 4 - Brief description of the development environments identified in the data collection .............................. 25 5 - Solutions for additional use cases within evidence management .......................................................... 26 6 - Additional features of tools ................................................................................................................ 28 www.efsa.europa.eu/publications 6 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table 7 - Gaps and Recommendations per each dimension ............................................................................... 33 8 - Comprehensive overview of cooperation models ................................................................................. 36 9 - Development Levers ......................................................................................................................... 38 10 - Institutional organisation prioritisation ............................................................................................. 43 11 - Collaboration Opportunities per Prioritised Use Case ......................................................................... 45 12 - Mapping between AI Act requirements and recommendations and actions for EFSA ............................ 64 13 - Map of the HLEG’s Requirement with short description ...................................................................... 66 14 - List of ETAPAS principles with short description ................................................................................ 67 15 - Example of organisational governance ............................................................................................. 70 16 – AI Adoption Scenarios and connected Prioritised Use Cases of the Evidence Management .................. 77 17 - Prioritisation of Horizontal and Vertical Recommendation .................................................................. 90 18 - Roadmap for actions (Horizontal & Vertical Recommendations) ......................................................... 92 19 - Expected budget and time scale of the Horizontal Recommendations ................................................. 94 20 - Expected budget for Vertical Recommendations ................................................................................ 94 21 - Expected budget, time scale and man days for Programme Management ........................................... 95 22 - Expected Total Budget per Horizontal and Vertical Recommendations ................................................ 95 23 - List of EFSA stakeholders' categories and sub-categories ................................................................... 97 24 - Stakeholder value map .................................................................................................................. 105 25 - Concepts to integrate into Communication messages ...................................................................... 106 26 - Engagement plan.......................................................................................................................... 110 27 - Collaboration models and related tools ........................................................................................... 117 28 - Focus on synergies with other roadmaps ........................................................................................ 118 www.efsa.europa.eu/publications 7 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence List of acronyms Acronym Definition ADRA AI, Data and Robotics Association AESGP Association of the European Self-Care Industry AI Artificial Intelligence ALTAI Assessment List of Trustworthy Artificial Intelligence AMU Assessment and Methodological Unit ANEC European consumer voice in standardisation API Application Programming Interface ArcGIS Geospatial Information Management System ASySD Automated Systematic Search Deduplicator AWS Amazon Web Services BEUC European Consumer Organisation CDT Translation Centre for the Bodies of the European Union CEF2 The Connecting Europe Facility 2021-2027 CoC Code of Conduct CNR National Research Council CPVO Community Plant Variety Office CV Curriculum Vitae DATA Evidence Management Unit (EFSA) DCF Data Collection Framework DEP Digital European Programme DevOps Development and Operations DG Directorate General DG Data Governance DIH Digital Innovation Hubs DOI Digital Object Identifier DTs Disruptive Technologies EC European Commission ECDC European Centre for Disease Prevention and Control ECHA European Chemicals Agency ECOTOX Ecotoxicology Knowledgebase EEA European Environment Agency EFA European Free Alliance EMA European Medicines Agency EP European Parliament EPA United States Environmental Protection Agency www.efsa.europa.eu/publications 8 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Acronym Definition EPHA European Public Health Alliance EFSA European Food Security Authority EKE Expert Knowledge Elicitation ENISA European Union Agency for Cybersecurity ETAPAS Ethical Technology Adoption in Public Administration Services, Horizon 2020 project ETL Extract Transform and Load EU European Union EUIPO European Union Intellectual Property Office FAO Food and Agriculture Organization FVE Federation of Veterinarians of Europe GO Gene Ontology H2020 Horizon 2020 programme HLEG High-Level Expert Group ICPEN International Consumer Protection and Enforcement Network ICT Information and Communications Technology IUCLID International Uniform Chemical Information Database JRC Joint Research Centre KJC Knowledge Junction Community MCDA Multi-Criteria Decision Analysis ML Machine Learning MoU Memorandum of Understanding NAMs New Approach Methodologies NCA National Competent Authorities NCBO National Center for Biomedical Ontology NER Named Entity Recognition NGO Non-Governmental Organisation NLP Natural Language Processing OBO Open Biological and Biomedical Ontology OCR Optical Character Recognition OECD Organisation for Economic Co-operation and Development PA Public Administration PARC European partnership for the assessment of risks from chemicals, Horizon Europe PERA Partnership for Environmental Risk Assessment PCP Pre-commercial Procurement PoC Proof of Concept PPI Public Procurement of Innovative Solutions www.efsa.europa.eu/publications 9 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Acronym Definition R&D Research & Development RIS Research Information Systems RM Roadmap ROI Return on Investment SaaS Software-as-a-Service SDU Smart Document Understanding SEPAL System for Earth Observation Data Access, Processing, and Analysis for Land Monitoring SLR Systematic Literature Review SMEs Small and Medium Enterprises SNAPP Science for Nature and People Partnership SR Systematic Reviews SRA Quantitative security risk assessment SQL Structured Query Language TIM Tools for Innovation Monitoring TRL Technology readiness level UCL University College London VKM The Norwegian Scientific Committee for Food and Environment WHO World Health Organization www.efsa.europa.eu/publications 10 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence 1. Introduction This report is an outcome of an EFSA procurement titled “Development of a Roadmap for Actions on artificial intelligence for evidence management in risk assessment (specific contract No 35 –RC 5 implementing Framework Contract No OC/EFSA/FIN/2019/01), awarded by the European Food Safety Authority to PwC EU Services EESV and the subcontractor Intellera Consulting Srl. This report is the final document of the “ Roadmap for actions on artificial intelligence for evidence management in risk assessment” project (in brief “AI-RM”), conducted by PwC EU Services EESV and its subcontractor Intellera Consulting SLR for the European Food Safety Authority (EFSA). It presents all the activities performed, together with the elaboration and the analysis of the main findings collected during the project. The report is structured as follows: • Chapter 1 “Introduction” provides an overview of the content of each chapter of the report. • Chapter 2 “Background” includes the context in which the project was conducted, with a focus on the European policy framework and on EFSA’s work on AI development. In addition, it presents the high-level methodology and the problem formulation used to define the evidence management use cases that guided the subsequent project activities. • Chapter 3 “State of the art and analysis” provides an overview of the state of the art of AI solutions based on the results that emerged from the various data collection activities. In addition, a dedicated section explores the potential challenges and blockers related to the adoption of AI, followed by the identification and evaluation of collaboration opportunities with key players in the market. • Chapter 4 “Recommendations and prioritisation” outlines the roadmap’s recommendations (horizontal recommendations and vertical recommendations) together with the prioritisation of working areas based on the outcomes of the first activities of the project. • Chapter 5 “Communication and engagement opportunities” outlines potential communication and engagement opportunities in the field of AI approaches to evidence management of risk assessments, taking into consideration specific needs and concerns of EFSA’s stakeholders on the adoption of AI. • Chapter 6 “Conclusions” provides final considerations on the main outcomes of the project based on the content provided in the chapters previously outlined. 2. Background This chapter focuses on the context in which the project was conducted, with a focus on the European policy framework and on EFSA’s work on AI development. In addition, it presents the specific objectives of the project, the high-level methodology and the problem formulation used to define the evidence management use cases that guided the subsequent project activities. The chapter is structured as follows: • Section 2.1 “Policy context” provides an overview of the European policy context that has been evolving with regard to AI, taking into consideration regulatory actions from the European Commission. • Section 2.2 “EFSA’s specific objectives” provides context concerning the specific objectives to be achieved through this project. • Section 2.3 “High level project phases” provides an overview of the various phases of the project. • Section 2.4 “Problem formulation” provides an overview of the main activities conducted in order to frame the project. www.efsa.europa.eu/publications 11 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence 2.1. Policy Context Over the last years, growing computing power and increased data availability, have led to a significant rise in the strategic importance of artificial intelligence (AI) in the European scenario. AI is revolutionising industry and society worldwide, allowing for significant changes and opportunities, while also presenting new challenges. Indeed, AI is considered a “disruptive technology”, with the potential to change the way products and services are created and delivered across a wide range of industries. Using AI technologies in the public sector can promote openness, accountability, quality and trustworthiness of given services, as well as saving costs and time, but it also entails specific risks and challenges. In this context, the goal of the European Union (EU) is to make the EU a world-class hub for AI, while ensuring that AI is human-centric, trustworthy and grounded in European values and fundamental rights. To this end, in April 2018, in its communication “Artificial Intelligence for Europe”, the European Commission (EC) reaffirmed the role of AI as one of the most strategic technologies of the XXI century, representing a key driver of economic growth. Since then, the development of the AI market – and ensuing regulatory needs – has always been high on the EC’s agenda, as demonstrated by the increasing amount of financial investments and policy documents adopted. Figure 1 below highlights the most relevant policy and regulatory initiatives launched by the European Commission since 2018. Figure 1 – EU’s approach to AI. As highlighted in the previous figure, the EC published the proposal for an “AI Regulation” on 21 April 20212, also called the AI Act, which aims to foster the development and application of AI in the internal market, while addressing - and mitigating - the risks generated by specific uses of AI through a set of complementary, proportionate and flexible rules (legal framework for trustworthy AI 2). The regulation’s primary objective is to “ensure the proper functioning of the internal market by setting harmonised rules, in particular on the development, placing on the Union market and use of products and services making use of AI technologies, or provided as stand-alone AI systems”.3 In practice, a risk-based approach is put forward to address and manage risks associated with AI systems. Specifically, risks are classified according to four clusters: 2 European Commission, Proposal for a Regulation Of The European Parliament And Of The Council Laying Down Harmonised Rules On Artificial Intelligence (Artificial Intelligence Act) And Amending Certain Union Legislative Acts, COM (2021) 206 final. www.efsa.europa.eu/publications 12 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence ● ● ● ● Unacceptable risk: All AI systems that are banned as they are considered a clear threat to the safety, livelihoods and rights of people, e.g., from social scoring by governments to toys using voice assistance that encourages dangerous behaviour; High-risk: AI systems employed in critical sectors (e.g., transport, education, migration) and those that are subject to strict obligations (i.e. in terms of quality, functionalities, transparency, security and robustness) before they can be put on the market; Limited risk: AI systems that are permitted but subject to specific transparency obligations (e.g. chatbots); Minimal risk: AI systems with very low or no risk that are allowed without restrictions (e.g. AI-enabled video games, spam filters). The approach is aimed at preventing "high-risk" applications from entering the market. High-risk applications are those relating to employment, transport, and law enforcement, for example. These systems must meet five requirements: data quality, bias, explainability, human oversight, security, and accuracy. High-risk systems must be registered with the EU and re-assessed if the system changes substantially. Applications that manipulate human behaviour to evade users' free will and government social scoring are totally prohibited as they pose an intolerable danger. With rare exclusions, biometric surveillance in public settings is likewise forbidden . In addition, the legal framework aims to reduce administrative and financial barriers for businesses. In particular, it fosters innovation through sandboxing programmes, aimed mostly at small and medium-sized enterprises (SMEs). The proposal also includes the creation of the "European Artificial Intelligence Board". In order to prepare EFSA to the disruptive impact that the regulation is likely to have on the way AI is used and managed, in the context of this project, the trustworthiness of AI has been taken into consideration as a fundamental aspect, therefore explored both during data collection activities and during the subsequent development of recommendations. 2.2. EFSA's specific objectives EFSA's interest in Artificial Intelligence is in line with the EC's agenda. In the last years, together with other Agencies, EFSA has launched the project “Joining forces at EU level on the implementation of Artificial Intelligence”, in order to explore the benefits of AI technologies and develop a European virtual community around AI. The project provided an overview on AI Trend Analysis, AI Maturity Assessment and the AI strategic roadmap. Moreover, analysing EFSA internal processes, it emerged that one of the most relevant uses of AI should be applied to Evidence Management Forecasting, upon which EFSA developed other interesting AI techniques. This project was conceived also as an effect of the “Transparency Regulation”3 published by the EU, that in its Article 32d states that, in exceptional circumstances of controversy or conflicting results, scientific studies may be commissioned, with the objective of verifying evidence used in its risk assessment process, and having eventually a wider scope than the evidence subject to verification. To this end, EFSA´s grant and procurement budget embed verification studies on an annual basis from 2021 onwards, and in the absence of specific requests under Article 32d, the Agency will dedicate part of its grants and procurement budget for the purpose of preparedness for verification studies. In addition, in line with its consideration on the future of science 4 and its 2027 strategic development 5, in 2021 EFSA prioritised some scientific themes requiring multi-annual cooperative arrangements with Member States, EU Agencies and where relevant, international partners. For each of them, a roadmap for action is being developed, 3 Regulation (EU) 2019/1381 of the European Parliament and of the Council of 20 June 2019 on the transparency and sustainability of the EU risk assessment in the food chain and amending Regulations (EC) No 178/2002, (EC) No 1829/2003, (EC) No 1831/2003, (EC) No 2065/2003, (EC) No 1935/2004, (EC) No 1331/2008, (EC) No 1107/2009, (EU) 2015/2283 and Directive 2001/18/EC. OJ L 231, 6.9.2019, p. 1-28. 4 Bronzwaer, S., Kass, G., Robinson, T., Tazarona, J., Verhagen, H., Verloo, D., Vrobs, D., & Hugas. M (2019). Food Safety Regulatory Research Needs 2030; EFSA Joural 5 https://www.efsa.europa.eu/sites/default/files/event/mb-82/mb191218-7-p.pdf www.efsa.europa.eu/publications 13 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence in order to define specific projects or studies to carry out in the following years in order to attain the strategic objectives. One of these scientific themes concerns the use of AI in the evidence management process of risk assessment. This roadmap provides recommendations for future multi-annual, multi-partner studies or projects in the field of AI applications to the evidence management process, by building on EFSA’s vision and supporting its preparedness for future risk assessments, while preventing possible discrepancies on sensitive matters6. The roadmap for action provides: • • • A full understanding of ongoing activities, knowledge gaps, and societal interests and concerns, as well as collaboration opportunities and potential partners, A set of recommendations for the adoption of AI in the evidence management process, with a proposed timeline and budget estimate, A communication and engagement plan in view of the adoption of AI. By developing an approach for the implementation of AI methods in the evidence management phase of risk assessment, EFSA will achieve its vision included in the Theme paper on AI in risk assessment 7 by: ● increasing accessibility of the body of evidence, ● increasing breadth of the body of evidence, ● enhancing trustworthiness of the subsequent risk assessment process, applying human centric artificial intelligence in close co-existence with human expertise. 2.3. High level project phases The roadmap was developed in 8 months, and was composed of three subsequent phases: • The Inception phase which aims at refining the scope, methodology and workplan of the entire project, based on the Technical Offer and the input gathered at the Kick-off meeting by the client. • The Execution phase which foresees the data collection activities, data analysis and definition of interim output needed to draft the Roadmap Report. • The Roadmap definition phase which includes the synthesis activities of the interim output and the design of a communication and engagement strategy in the field of AI. Figure 2 below shows the main activities carried out for each phase and related duration : 6 See also Regulation (EC) No 178/2002 of the European Parliament and of the Council of 28 January 2002 laying down the general principles and requirements of food law, establishing the European Food Safety Authority and laying down procedures in matters of food safety. OJ L 31, 1.2.2002, p. 1–24.; Regulation (EU) 2019/1381 of the European Parliament and of the Council of 20 June 2019 on the transparency and sustainability of the EU risk assessment in the food chain and amending Regulations (EC) No 178/2002, (EC) No 1829/2003, (EC) No 1831/2003, (EC) No 2065/2003, (EC) No 1935/2004, (EC) No 1331/2008, (EC) No 1107/2009, (EU) 2015/2283 and Directive 2001/18/EC. OJ L 231, 6.9.2019, p. 1-28. 7 European Food Safety Authority, Theme (concept) paper – on Artificial intelligence in risk assessment (2021) www.efsa.europa.eu/publications 14 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 2 - Main activities of the project. 2.4. Problem formulation In order to clarify the scope of the roadmap and AI integration potential within the process, the following activities have been performed: • High level mapping of the evidence management process, performed on the basis of the available information and the inputs of the EFSA staff; • The identification of potential use cases, i.e. specific process steps or tasks which can be improved by AI tools/approaches; • The assessment and prioritisation of use cases, aimed at defining a shortlist of key use cases to guide the following project activities. The prioritisation was an iterative procedure, based on the different inputs collected during project execution. 2.4.1. High-level mapping of the evidence management process in EFSA EFSA is the risk assessor, evaluating risks associated with the food chain. EFSA collects and analyses existing research data to provide independent scientific advice, supporting the decision-making of risk managers. In addition, by collecting and analysing available data, it communicates existing and emerging risks associated with the food chain. Therefore, EFSA's evidence management process is at the core of the evidence-based risk assessment. Specifically, evidence management is a structured process of collection and analysis of various streams of evidence (e.g., scientific data, expertise literature, best practices, etc.) generated through multiple methods and reported by multiple sources. It follows a specific and documented process with clearly defined phases managed by different Units in EFSA – including AMU (Assessment and Methodological unit) and DATA (Integrated Data unit). The highlevel evidence management process steps are the following: 1. Evidence Collection: systematically gathering and validating the required streams of data through multiple means and from various reliable sources (relevant data providers e.g. (EU Member States organisations’ databases, literature papers and from experts’ knowledge in the field) 2. Evidence Appraisal: assessing the quality, consistency and complementarity of all the data collected from the different sources; 3. Evidence Synthesis and Integration: combining, summarising and finding key patterns within data through the performance of multiple typologies of data analysis (e.g., statistical analysis, meta-analysis, meta-regression and dose-response modelling) and consequently interpreting the results; 4. Evidence Visualisation and dissemination: the production of customised data visualisations - in the form of reports, tables and graphs -and publication of evidence in the EFSA curated and open repository for the exchange of evidence and supporting materials used in food and feed safety risk assessments (Knowledge Junction Community) www.efsa.europa.eu/publications 15 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 3 - Evidence management process phases. The succession of these phases provides a fit for purpose and transparent risk assessment. These macro-phases of evidence management are composed of single processes (see table 1), which in turn can be broken-down into steps. For each process, key considerations on the current degree of automation and technology architecture were put forward to identify relevant pain points and risks within the process, thus shedding light on potential applications of AI technologies. Table 1 - Risk Assessment Workflow and Evidence Management Process. Macro phase Process Process description Key Considerations Data collection from Member States, data holders, and other stakeholders It consists in receiving, processing and validating structured data through the DCF (Data Collection Framework, i.e. EFSA Online Tool) from multiple stakeholders Process with low level of automation Systematic Literature Review It consists of the methodological extraction, appraisal, analysis and reporting with respect to sources within the existing literature (e.g. peer-reviewed articles) Time-consuming process, with a low level of automation and subject to human error risk and bias Expert Knowledge Elicitation It consists of eliciting relevant knowledge from one or more experts in the field to complement the previous processes and bridge limited, absent or conflicting empirical evidence Complexity in finding relevant experts’ contacts with the required competences and knowledge and variety in their selection Dossier assessment It includes the submission of dossiers to EFSA as part of the Application Management risk assessment process Continuous improvement of digitalisation due to the introduction of the Appian platform Public consultations It consists in the collection and management of comments from key stakeholders in different cases (e.g. comments on the list of intended studies and their design for application renewal) Highly labour-intensive process with a low level of automation B Evidence Appraisal Advice on Evidence Appraisal It includes the provision of advice and quality evaluation of all types of integrated evidence streams submitted or selected by EFSA Time-consuming process execution with a rather low level of automatisation and prone to possible human errors C Evidence synthesis and integration Evidence analysis It consists of inspecting, cleansing, transforming, modelling and interpreting data with the goal of discovering useful information, informing conclusions, and supporting decision-making. Process with low degree of technology development/maturity and low level of automation A Evidence Collection www.efsa.europa.eu/publications 16 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Macro phase Process Process description Key Considerations Uncertainty Analysis It consists in the systematic identification and characterisation of uncertainty and limitations in questions of interest and/or quantities of interest in a scientific assessment, and evaluating their implications for scientific conclusions. Process variant with low degree of technology development/maturity and low level of automation Production of Reports and Data Visualisation It consists in producing relevant data visualisations and reports through statistical tools Lack of automation in the production of visual reports concerning data collected from the SLRs and the public consultation. Publication on the Knowledge Junction Community It consists of the publication of curated supporting material/evidence (e.g. annexes of scientific outputs containing raw data, specific documents of external organisations during the public consultations of a scientific output) on the Knowledge Junction Community and the maintenance of the latter in terms of requirements, permissions and metadata Lack of automation in the maintenance of the KJC repository D Evidence visualisation and dissemination As previously mentioned, this process plays a crucial role within EFSA’s risk assessment, but it is a resource-intensive activity where the amount of available information is limited both by the accessibility of the evidence itself and by the available resources to undergo the different phases of the process. For this reason, AI has been considered a key technology to improve the process in terms of costs, time and quality of the output. Artificial intelligence (AI) approaches have a strong potential in reducing human-bias in the process and reducing the need to initiate verification studies (as per Article 32d of Regulation (EU) 2019/1381). In case a verification study is required, AI approaches will facilitate the integration of the new evidence generated by the verification study. 2.4.2. Identification of the use cases Through an internal consultation with EFSA, a clear understanding of the agency’s processes was achieved. The information collected made it possible to identify an extensive list of use cases, i.e., specific tasks within the evidence management process, which can potentially be automated through AI. In addition, data collection activities made it possible to gather further information on the potential of the use cases in terms of impact and feasibility. The output of these activities can be summarised in the extensive list of use cases shown in figure 4 below. www.efsa.europa.eu/publications 17 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 4 – Evidence management process phases and use cases. 2.4.3. Prioritisation of use cases In order to guide the project activities, we performed a prioritisation exercise on the long list of use cases. This prioritisation was an iterative procedure that occurred in multiple rounds, based on the additional inputs collected during the project. We set out below the final output of this activity. Our approach was based on a Multi-Criteria Decision Analysis, or MCDA, a method that is commonly applied to face complex decisions and to solve problems involving a choice among alternatives. The approach had the advantage of allowing us to focus on what is important in a logical, coherent, and proportionate way. To apply MCDA for prioritising the AI use cases, the following steps were performed: • Criteria Selection for the purpose of dividing the decision into smaller, more understandable parts in the form of criteria tailored to the subject at hand. The approach for selecting the relevant criteria was firstly based on a standard impact/potential for adoption matrix approach. Within these broad categories, specific criteria were selected to reflect EFSA's vision, stakeholder's opinions as registered during the working sessions and a preliminary feasibility analysis of each use case's functionality. As a result of this analysis, the dimensions and respective criteria used in the prioritisation exercise were the following: Dimension 1: Impact, which included the following three criteria: – Time Efficiency/Effort Reduction: expressed as the reduced time duration and effort (i.e., hours workers spent) to perform a process or a task; – Quality Improvement: expressed as an enhanced process performance (e.g. via an increase in the accessibility of evidence, enhanced analysis capacity, etc) and delivery of better outputs in terms of content and quality of conclusions; – Reduction of human bias: expressed as a lower risk of human errors and deviation from norm or objectivity in any selection, appraisal and decision-making steps. Dimension 2: Potential for adoption, which included the following two criteria: – Acceptance: expressed as the potential for acceptance of the integration of AI in the proposed use case by the relevant users and stakeholders within the organisation; www.efsa.europa.eu/publications 18 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence – Feasibility: expressed as the degree of readiness in a functional level of the proposed use cases and the potential integration of the AI solution into the current operations (e.g. in terms of process framework). • Scoring of each use case according to each criterion. The approach was used to provide a score for each criterion for each individual use case. The scoring approach assigned a score on a scale of 0-3, where 0 equals no impact/potential, 1 equals low impact/potential, 2 equals medium impact/potential and 3 equals high impact/potential. For the Impact dimension criteria (time efficiency/quality improvement/reduction in bias) we based our scoring on our understanding of the issues faced by EFSA at each stage of the evidence management process, through the discussions with the various teams and stakeholders, and considering the improvements that could be made via the use of AI. For the Potential for Adoption dimension criteria (acceptance/feasibility) we based our score on the discussions with EFSA’s teams and stakeholders, and also our own knowledge and understanding of the current state-of-the-art in AI-enabled solutions for providing satisfactory solutions to each use case. • Integrating the scored criteria though a weighted average approach to produce a meaningful solution and identify the final ranking. In order to reach an integrated scoring, we used a weighted average approach, where the Impact and Potential for Adoption dimensions had an equal weight. The choice of weights for individual criteria was based on our understanding of EFSA’s general purpose and objectives, taking into consideration the opinion of EFSA’s internal stakeholders and our own understanding of the individual use case complexities, as well as the current readiness level of AI applications on the market. The weighting system has been tested through a sensibility analysis, and then validated by the project team. Figure 5 – MCDA criteria for use case prioritisation The output of the revised prioritisation is represented graphically in figure 6 below. The use cases highlighted in green have the highest levels of Impact and Potential for Adoption, thus falling into the prioritised use cases list (also listed in table 2 below). www.efsa.europa.eu/publications 19 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 6 - Revised prioritisation of the use cases. Table 2 - Top 10 prioritised use-cases. Use Case Impact Potential for Adoption Prioritisation Score SLR Keywords Identification 2.5 3 5.5 2. SLR Literature Findings Clustering 2.5 2.75 5.25 3. Expert Selection 3 2.25 5.25 4. Expert Pool Identification 2.75 2 4.75 5. Comparable Appraisals Detection 2.5 2.25 4.75 6. SLR Relevant Abstract Screening 1.75 3 4.75 7. Automatic Text Summarisation 2.25 2.25 4.5 8. Quality Improvement of the Scientific Output 2 2.25 4.25 9. SLR Literature De-duplication 2 2.25 4.25 10. Data Collection Terminology Assessment 2.25 2 4.25 # Use Case 1. In conclusion, according to our understanding of the state of the art combined with EFSA’s needs, the prioritised use cases are the most feasible and impactful for the adoption of AI in EFSA’s evidence management process. Therefore, prioritised use cases have been considered to develop the set of “vertical actions” for the roadmap (see chapter 4). 3. State of the art and analysis This chapter focuses on the current state of the art of AI in evidence management, taking into consideration existing solutions and initiatives identified during the project. The main findings are analysed, providing information related to potential challenges and blockers to be encountered when applying AI. Finally, potential cooperation and collaboration opportunities, identified during data collection activities, are provided. The chapter is structured as follows: • Section 3.1 “Market overview, map of relevant activities and further areas” provides a description and some key results regarding the multiple AI initiatives and solutions and related organisations investigated. These were identified with various activities performed during the project (desk research, www.efsa.europa.eu/publications 20 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence survey, interviews, workshop), covering the prioritised evidence management use cases process. It also focuses on further AI areas that fall outside of the scope of evidence management • Section 3.2 “Challenges and blockers” introduces general challenges and blockers for adopting AI in a public organisation, together with specific challenges that EFSA would face for the adoption of AI in its evidence management process in the form of a gap analysis of the AI and data maturity assessment of EFSA’s ecosystem. • Section 3.3 “Cooperation and collaboration opportunities” provides an identification of the phases of the end-to-end AI lifecycle, from strategy to operation, together with possible collaboration/cooperation models to be adopted with multiple stakeholders. The link between AI lifecycle phases and collaboration models will be further investigated in section 5.4. 3.1. Market overview, map of relevant activities and further areas The data collection phase was performed through multiple methods – desk research, literature review, stakeholders’ survey, interviews with relevant stakeholders and workshop. Subsequently, the data collected during the project has been analysed in order to map primarily key initiatives, tools and the respective organisations in the area of AI approaches in evidence management, and secondly, to identify additional information such as the policy background that is still relevant for the development of the roadmap. In addition, the same data collection activities made it possible to further investigate AI applications that fall outside of the scope of the evidence management. 3.1.1. Initiatives The term “initiatives” refers to the activities of organisations – especially institutions – related to the use of AI in the evidence management process (e.g., organisation of a webinar, pilot projects, tender procedures, development of an internal algorithm). This was one of the most important topics explored through the data collection activities, as it can be useful to understand whether there are other organisations with similar needs, and what has been the impact, challenges and lessons learned from these experiences. In general, only 8 initiatives related to the use of AI in the evidence management process have been identified (see Table 3 below). Some other projects on the use of AI have been identified as well, but they fall outside of the scope of this project. For example, we found some projects and roadmaps of AI related to the agricultural sector (e.g. FAO introduced an Advanced Land Monitoring system used for earth observations, data access, processing & analysis for land monitoring), and AI applications for the analysis of social media unstructured data (e.g. ECDC has launched a free, open-source interactive tool called “Epitweetr” to automatise a more detailed early detection of public health threats using Twitter data using the R programming language, allowing them to automatically monitor trends of tweets by time, place and topic, thereby detecting signals such as an unusual increase in the number of tweets). It is worth mentioning that, even though some institutions have projects related to evidence management, on the basis of the evidence gathered, no detailed strategies, roadmaps or action plans on this topic have been developed by them. The following list of initiatives (Table 3) was drafted by combining the information collected by the different data collection activities, including the literature review, desk research, survey and interviews. This table provides an overview of the initiatives developed by institutional stakeholders (EU Agencies and Institutional, International organisations and National competent authorities). Table 3 - Extensive list of initiatives. Organisation Short description of the initiative CDT The organisation is currently working at adopting an AI-based tool to automatically summarise long texts, aiming to significantly reduce costs on single translations and decrease the number of pages that need to be translated, ultimately enabling a higher number of documents to be translated www.efsa.europa.eu/publications 21 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Organisation Short description of the initiative The organisation has adopted AI for text extraction, where an open-source solution able to read letters and extract relevant information is used. EUIPO’s AI model is internally developed. EUIPO The organisation, in collaboration with the IE University, presented through a webinar a detail review of AI’s developments, both conceptually and in terms of its relevance to specific business situations. Among the most relevant applications, AI can be employed for text extraction (e.g., more specifically anomaly detection, automated table extraction), analysis and summarisation and natural language generation. The organisation is leading on two recent pilot projects, each focusing on one use case of evidence management, mainly Text data structuring (data extraction and structuring from summary product characteristics) and Keywords/metadata identification. The organisation has produced a report named ‘Evolving Data-Driven Regulation’ that outlines several recommendations on leveraging Artificial Intelligence (AI) and new digital technologies, to improve efficiency in processes and increase insights into data. These recommendations aimed to assist regulators and stakeholders in capturing the opportunity for data-driven, evidence-based, rigorous decision-making that will underlie the development, approval, and on-market safety and effectiveness monitoring of medicines, in a fast-expanding data and analytics world EMA ENISA Member of the Blueprint cooperation framework (EC), ENISA has opened a Tender procedure to further develop the “Open Cyber Security Situation Awareness Machine,” an Open-Source Intelligence tool that aims to optimise the production of cybersecurity situation awareness reports, by applying Natural Language Processing (NLP) and Machine Learning (ML) techniques for the automatic classification and generation of documents. OECD The organisation has recently launched a workshop called “AI and the productivity of science”, which was pointed out by a Senior Policy Analyst that we interviewed. The workshop addresses the critically important issue of the rate of scientific progress, whether this is stagnating, as recently argued by a number of scholars, and how AI could raise the pace of progress in science and discovery. The workshop brings together technical and policy experts to examine the evidence on a purported productivity decline in science, as well as the ways that AI is currently used across different fields of science – from neuroscience to materials science - and across all stages in the scientific process. United States Environmental Protection Agency (US EPA) The US Environmental Protection Agency (EPA)'s etc has developed a novel approach to article retrieval in the PubMed Abstract Sifter. This approach has been implemented and made publicly available in the Dashboard and in a Microsoft Excel version. The Dashboard version is a convenient and powerful way to find information quickly and efficiently about a chemical of interest. For more in-depth literature tasks such as reviews, the Excel version of the Abstract Sifter offers additional functionality. Results from a PubMed search are imported directing into an Excel sheet, where the end-user can then use a novel “sifter” methodology for quick, agile relevance ranking of articles. The tool also enables article triage capabilities through easy tagging and noting functionality. Triaged citations can be exported to external software such as reference management tools. The Abstract Sifter can also provide a high-level view of a corpus of literature for a defined set of entities such as chemicals. 3.1.2. AI solutions and respective organisations This paragraph provides an overview of the information related to the identified solutions through the different data collection activities performed since the beginning of the project, namely desk research, literature review, survey, interviews and workshop, as shown in Figure 7. www.efsa.europa.eu/publications 22 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 7 - Data collection activities. A total of 81 solutions were identified, which were then classified according to use cases, relevance and level of readiness8. Out of the 81 solutions identified, 35 are currently associated with a high readiness level, 40 with a medium readiness level, and 7 with a low readiness level. It is worth mentioning that some solutions cover more than one use case: out of the 81 solutions, 49 cover only one prioritised use case, 13 cover 2 prioritised use cases, and 5 cover 3 prioritised use cases (figure 8). 14 tools were associated to some non-prioritised evidence management use cases (e.g., Public Consultation Automation, Confidentiality Dossier Assessment, Citations Screening, Information Consolidation). Figure 8 - Use case relevance. In terms of use cases association, 23 solutions were related to Literature Clustering, 10 to Literature deduplication, 15 to Keywords identification, 13 to Experts selection, 4 to Expert Pool Identification, 9 to Improvement of scientific output, 4 to Comparable appraisal detection, 8 to Terminology assessment, 12 to Text summarisation, and 4 to Abstract screening (Figure 9). The following graph (Figure 9) depicts the relationship between the solutions, use cases and readiness level. 8 Level of readiness definitions: Low level: Refers to tools/solutions that are in a research stage Medium level: Refers to tools/solutions that are in a development stage High level: Refers to tools/solutions that are in a deployment or commercialisation stage www.efsa.europa.eu/publications 23 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 9 - Number of solutions per use case and readiness level. Overall, the Evidence Collection phase within the Evidence Management process represents the area with the greatest room for AI-based automation and improvement possibilities, leading to multiple potential benefits such as time and effort reduction and quality improvement. Within the Evidence Collection phase, Data Collection from multiple stakeholders, SLR, and Expert Knowledge Elicitation are the areas that record the highest number of developed or in development solutions, according to our analysis. Furthermore, most solutions explored through the different data collection activities for this project can be positioned into an advanced or medium phase of development (high or medium readiness), and this might open opportunities for collaboration with key stakeholders (i.e., for the co-development of a tool, or co-execution of research), resulting in possible purchases (in case of commercialised tools), and allowing the sharing of knowledge and lessons learnt. The following image (Figure 10) shows a summary of the main key insights, including the use case maturity & relevance, the type of stakeholder which developed the solutions identified, and their locations. Figure 10 - Summary of key insights. www.efsa.europa.eu/publications 24 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence 3.1.3. AI Development environments and services This section presents an overview of four cloud-based AI development services identified in the data collection phase of the project, which can provide EFSA with a suitable suite of ready-made data preparation pipelines, ready-made AI-based models and services and code development modules to enable the end-to-end development of AI solutions, and in particular text analytics/natural language processing (NLP), to address the use cases relevant to evidence management in risk assessment, and also EFSA's wider needs and plans with regard to the adoption of AI. The identified services represent market leaders in NLP services and development environments: Microsoft Azure, IBM, Amazon Web Services and Google cloud. Table 4 below provides a short description of the development environments identified. Table 4 - Brief description of the development environments identified in the data collection. Service Brief Description Azure Cognitive Services Azure Cognitive Services are cloud-based services available to help users build cognitive intelligence which can be customised. The development environment allows the inclusion of cognitive features in AI applications without having artificial intelligence (AI) or data science skills. Azure Cognitive Service for Language provides several Natural Language Processing (NLP) systems, including Text Analytics, which is a collection of features from Cognitive Service for Language that extract, classify and understand text within documents (relevant to several use cases in evidence management, especially within the SLR activities). In addition, the service includes a text summarisation module. IBM Watson IBM Watson Natural Language Understanding is a service provided by IBM cloud which uses deep learning to extract meaning and metadata from unstructured text data. It offers services around text analytics to extract categories, classify text and extract trends from text, all of relevance to several of EFSA's use cases in evidence management. In addition, Watson Knowledge Studio allows users to train Watson to understand the language of a specific domain and extract customised insights. Amazon Web Services Amazon Comprehend is a natural language processing (NLP) service that uses machine learning to extract key phrases, entities and sentiment from text for further analysis. This particular service is, for the time-being, more geared towards customer service and front-office applications, and is less mature than the Microsoft and IBM solutions for use in a scientific context. Google Cloud Cloud Natural Language is the Google cloud service which helps users derive insights from unstructured text using Google machine learning. One application offered which is of relevance to EFSA is the custom entity extraction, which permits the identification of domain-specific entities within documents – many of which don’t appear in standard language models – without having to spend time or money on manual analysis. In addition to the ready-made models, Google provides the AutoML Natural Language service which allows users to train a custom machine learning model to classify documents, extract information, or understand the sentiment based on a specific domain. These services are relevant to the prioritised use cases since a classification model which analyses a document and returns a list of content categories would fulfil, for example, the SLR Literature Findings Clustering. 3.1.4. Overlaps with R&I programmes With regards to the EU R&I funding programmes (e.g. Horizon Europe, Digital Europe Programme), these can be helpful to identify whether other organisations do similar things, while exploring potential cooperation opportunities. Although, currently, there are no specific calls or partnerships on the application of AI in evidence management within the European ecosystem, there is some potential for future deployment of AI technology within organisations across different sectors. www.efsa.europa.eu/publications 25 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence For instance, it is worth mentioning that, in 2020, the EU Commission and ADRA 9 (AI, Data and Robotics Association) signed the European Partnership on artificial Intelligence, Data and Robotics 10, which covers the deployment of AI technology in a broad sense (without focusing specifically on the evidence management area). By 2030, the partnership aims to boost European competitiveness, societal well-being and environmental aspects to lead the world in researching, developing and deploying value-driven trustworthy AI, Data and Robotics based on fundamental European rights, principles and values. In addition, out of the scope of AI technology, some other relevant calls for EFSA have been designed in the 2021 Work Programme of Horizon Europe, such as, for instance: - HORIZON-HLTH-2021-ENVHLTH-03-01: European partnership for the assessment of risks from chemicals (PARC)11 - HORIZON-CL6-2021-FARM2FORK-01-16: Identification, assessment, and management of existing and emerging food safety issues12 In light of this, it can be assumed that potential calls and initiatives focusing on AI adoption in evidence management could be launched in the next few years, as a response to the above-mentioned calls and partnerships. Therefore, a constant monitoring action should be planned to track their development and progress. To conclude, in case of relevant calls for the use of AI in internal processes – and evidence management in particular - it would be important for EFSA to participate as an interested party in order to be fully aligned on the outcomes of the projects, to be involved within the community of stakeholders, and eventually to contribute to the development or testing activities and benefit from the projects’ results. 3.1.5. AI for additional use cases within the evidence management process Even though the data collection phase focused on the prioritised use cases of the evidence management process, and how the relevant tools can be implemented in one or more of these cases, additional solutions for the other use cases have been identified. These solutions are presented in this section (Table 5), as they can potentially be beneficial for EFSA and, if implemented, they could introduce additional innovation in the evidence management process. Table 5 - Solutions for additional use cases within evidence management. Other Use Case Anomaly Detection Readiness level13 Tool Tool Description Logicmelt Technologies SL Logimelt offers product development support services to technology companies that plan to use Artificial Intelligence in their products. Among other services, they offer custom design of Artificial Intelligence algorithms and optimised implementation of these algorithms in Edge Computing platforms. High 9 Composed of several stakeholders: BDVA, CLAIRE, ELLIS, EurAI and euRobotics https://ai-data-robotics-partnership.eu/ 11 European Commission; Horizon Europe Work Programme 2021-2022; Health. (European Commission Decision C(2021)9128 of 15 December 2021) 12 European Commission; Horizon Europe Work Programme 2021-2022; Food, Bioeconomy, Natural Resources, Agriculture and Environment. (European Commission Decision C(2021)9128 of 15 December 2021) 13 Level of readiness definitions: Low level: Refers to tools/solutions that are in a research stage Medium level: Refers to tools/solutions that are in a development stage High level: Refers to tools/solutions that are in an already commercialised stage 10 www.efsa.europa.eu/publications 26 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Other Use Case SLR Appraised Papers Data Extraction SLR Review Question Evaluation SLR Metaanalysis Execution Quantitative Data Synthesis and Integration Tool Description PhlexNeuron The PhlexNeuron regulatory automation platform provides a sophisticated machine learning framework that enables the extraction of specific data from unstructured data sources. High Syras Syras supports user to quickly review and categorise a large body of references. High WebPlotDigitizer WebPlotDigitizer is a web-based tool which extract numerical data from plot images or maps. High SWIFT-Review SWIFT-Review is an interactive workbench which provides numerous tools to assist with problem formulation and literature prioritisation. High EPPI - Reviewer EPPI-Reviewer is a web-based software program for managing and analysing data in literature reviews by identifying key terms. High Syras Syras support users to quickly review and categorise a large body of references. High Synthesis Synthesis allows user to automatically interact with network quantitative synthesis built from extracted data. High RevMan RevMan is software that supports the conduct of quantitative synthesis of data and the graphical presentation of results. High CMA CMA is a software package that allows the user to compute the effect size for each study automatically and create highresolution plots. High Neo4j AuraDB 3.1.6. Readiness level13 Tool Neo4j AuraDB is a fully managed cloud graph database service. Built to leverage relationships in data, AuraDB enables lightningfast queries for real-time analytics and insights. AuraDB is reliable, secure, and fully automated, enabling you to focus on building graph applications without worrying about database administration. High AI related to working areas outside of evidence management This section presents solutions that were identified as particularly relevant for different working areas outside of the evidence management process. Even though the literature review, desk research, survey and interviews focused on the prioritised use cases of the evidence management process, and how relevant tools can be implemented in one or more of these cases, additional functionalities and features have been highlighted. These features are presented www.efsa.europa.eu/publications 27 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence in this section (Table 6) as they can be beneficial for EFSA and, if implemented, also foster additional innovation in the risk assessment process. Table 6 - Additional features of tools. Tool Organisation Readiness level14 Description Main functionalities SGS Digicomply includes the tracking and analysis of different discussions in the media, and on social networks such as Twitter. High Sentiment analysis social media Google It provides entity analysis to find and label fields within a document—including emails, chat, and social media—and then sentiment analysis to understand customer opinions to find actionable product and UX insights. High Sentiment analysis AutoML Natural Language Google AutoML Natural Language enables you to build and deploy custom machine learning models that analyse documents, categorise them, identify entities within them, or assess attitudes within them. If you don't need a custom model solution, the Cloud Natural Language API provides content classification, entity and sentiment analysis, and more. High Sentiment analysis Crème Global Crème Global Crème Global includes a combination of a data and scientific models to quantify and understand risk, by implementing risk models in a web-based user-friendly interface. High Risk assessment analysis TIM (Tools for Innovation Monitoring) JRC It functions as a monitoring system for tracking the evolution of established and emerging technologies. Low Monitoring emerging technologies Medium Social media analysis Digicomply Natural Language AI Epitwitter ECDC Free, opensource interactive tool that helps with the automatised early detection of public health threats using Twitter data, and more in detail using the R programming language, allowing trends of tweets to be automatically monitored by time, place and topic, and detecting signals such as an unusual increase in the number of tweets. It is designed to support public health experts with the early detection of threats from infectious diseases, but can be extended to all hazards and other fields of study by modifying the topics and keywords. 14 Level of readiness definitions: Low level: Refers to tools/solutions that are in a research stage Medium level: Refers to tools/solutions that are in a development stage High level: Refers to tools/solutions that are in an already commercialised stage www.efsa.europa.eu/publications 28 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence These solutions, which fall outside of the evidence management working areas, should not be considered as relevant for reaching the vision of EFSA (i.e., the adoption of AI for increasing the body of evidence within the evidence management process by 2027). Nevertheless, they could be taken into account for other purposes such as automating time- and effort-intensive activities. 3.2. Challenges and blockers related to the adoption of AI in evidence management This sub-section provides insights concerning potential challenges and blockers on the basis of the information collected during the data collection phase of the project - including the analysis of the literature review, desk research, survey, interviews and workshop, as well as the information collected by internally consulting EFSA’s team for the implementation of AI solutions within the EFSA ecosystem. 3.2.1. General challenges and blockers The evidence gathered allowed us to identify four challenges and blockers related to the adoption of AI in a public organisation (See figure 11): - Adoption Challenge 1 – Impact on processes, governance and people: The first aspect includes challenges that the public organisation might encounter when considering the impact of AI adoption in terms of processes, governance, user acceptance and skills. This would require a more holistic organisational transformation for AI to be successfully adopted. Furthermore, specific processes and workflows might need to be re-designed to take into account the role of AI. In general, AI should not simply be viewed as the addition of automation and decision support within existing workflows, but instead it should be viewed as a transformational change that cannot only replace but also transform traditional ways of working. - Adoption Challenge 2 – Adopting solutions that can address specific problems: The second challenge concerns the need to identify AI solutions that fit specific needs and requirements. Organisations look for specific AI tools that can perform well on their own and that can provide answers to specific solutions, which are not always available in market. Organisations seeking specific solutions need to recognise that, for them to become truly AI-enabled, internal data science capabilities are needed, which can bring desired results in combination with the organisation’s domain expertise. - Adoption Challenge 3 – Ethical Risks: The third challenge considers issues related to the ethical risks of AI adoption, especially when used by public institutions, which must ensure the highest transparency and fairness standards. In fact, the adoption of trustworthy and human-centric AI should be a priority both internally and also in light of upcoming AI Regulation15. - Adoption Challenge 4 – Compliance: This challenge concerns the need for any use of data (within AI or not) to address compliance considerations around data protection (e.g., Art. 24 of EU GDPR "Responsibility of the controller"). These challenges - if not properly taken into consideration and addressed with dedicated measures - could hamper the efficacy and/or the feasibility of the adoption of AI in the evidence management process. For this reason, they are carefully considered during the roadmap development. Concerning technical challenges - related to the potential AI tools to be implemented – four categories have been identified as follows: i) Technical Challenge 1 – Readiness: The first technical challenge concerns the level of readiness of the AI solutions to be adopted. In fact, several tools are still at a premature/experimental stage of development and are not yet available in the market (i.e., not commercialised), therefore they would need a certain degree of development and/or customisation before being fully integrated. 15 European Commission, Proposal for a Regulation Of The European Parliament And Of The Council Laying Down Harmonised Rules On Artificial Intelligence (Artificial Intelligence Act) And Amending Certain Union Legislative Acts, COM (2021) 206 final. www.efsa.europa.eu/publications 29 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence ii) Technical Challenge 2 – Accessibility: The second technical challenge refers to the accessibility of the tool. Many tools, in fact, require advanced (e.g., programming language) technical skills in the case of customisable functionality. iii) Technical Challenge 3 – Applicability: The third technical challenge considers the applicability of the tool to specific domains of the organisation. Many solutions, in fact, are currently applicable to specific domains and, therefore, require customisation in order to be applied in different scientific domains. To address this, in addition to the need for developing internal data science capability (see Adoption Challenge 2 above), EFSA should invest in the development of subject-specific ontologies to customise commercially available or internally developed AI solutions. iv) Technical Challenge 4 - Autonomy: The last technical challenge relates to the dependence of some tools on human intervention. The level of maturity of commercial AI solutions still does not provide sufficient confidence in their use without a human-in-the-loop. The abovementioned technical challenges could hamper the feasibility and efficacy of the implementation of AI solutions in the different evidence management processes. Figure 11 - General Challenges. Finally, market readiness can be considered as a potential blocker for the adoption of AI solutions within evidence management in the next few years. The discontinuities and feasibility analysis highlighted those working areas (or use cases) where AI can already be adopted or implemented due to the presence of mature solutions in the market, and those where the AI market is still at an early stage . 3.2.2. EFSA’s challenges and blockers: Results of the Gap Analysis In order to identify technical or operational gaps that could hamper the feasibility or effectiveness of EFSA’s actions with regard to the application of AI in the evidence management process, a Gap Analysis was performed between EFSA current capabilities (AS-IS) and a target state which would support EFSA`s achievement of its vision (TO-BE) related to the adoption of AI tools. From a methodological point of view, an AI and data maturity assessment of EFSA has been performed, with emphasis on evidence management in risk assessment. The Gap Analysis consists of the following phases: www.efsa.europa.eu/publications 30 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence • Reviewing of the current state of EFSA’s capabilities, resources, procedures and structure for adopting the identified tools. • Describing the target state in terms of the technical requirements and skills required to customise and integrate the identified AI tools in the evidence management process of the risk assessment. • Based on the results of the previous activities, drafting a list of recommendations in order to guide the development of the full recommendations (see Chapter 4). 3.2.2.1. Current State of EFSA capabilities and future plans This section describes the current capabilities of EFSA's technical infrastructure (See Figure 12) with the aim of evaluating the adoption of AI solutions for evidence management in risk assessment. Figure 12 - Gap Analysis of the EFSA ecosystem. The following list provides information on the Technical Capabilities of EFSA staff & Procedures within EFSA: • Programming capabilities: SAS Analytics, SQL, Python and R, and handling XML and Json file formats. • APIs and data management capabilities: APIs including GET and POST methods, transmitting and parsing Json bodies, and data management and visualisation skills (MicroStrategy). • Data format procedure: data received from Member States are validated with rules implemented in SAS. • Scientific article storage procedure: scientific papers are downloaded from free and subscription-based databases. Currently, all papers are stored into a simple data management system. However, recently EFSA started to transfer them to Azure. In the list below, an overview of Hardware & Software Components, Structure & Requirements within EFSA is provided: • Cloud platform: EFSA has a contract to use Microsoft's Azure cloud platform and services. EFSA's vision is to take full advantage of the capabilities that this platform offers. • Software within the SLR working area: Currently EFSA is using the DistillerSR AI module for the partial automation of title and abstract screening phase of systematic literature reviews (SLRs), which approximately saves EFSA 60-70 % of the effort needed to accomplish this task. Even though EFSA is constantly looking for improvements of SLRs, replacing DistillerSR is not a desirable option as it would require a lot of effort www.efsa.europa.eu/publications 31 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence due to users’ resistance against the change. In addition, training on a new tool might be required. On the other hand, for full text screenings, no relevant tool is currently being used. Therefore, automating this process would be of a great benefit for EFSA. • Evidence repository: The Knowledge Junction is a curated, open repository for the exchange of evidence and supporting materials used in food safety risk assessments. • Software for statistical application: EFSA R4EU is a platform that is currently used to expose statistical applications internally and to the public by providing a user-friendly interface to perform sample-size calculations. The R4EU platform hosts a suite of different modules with different functionalities for modelling, commonly used in assessments by EFSA. Existing models include benchmark dose, multi-drug resistance analysis, risk assessment using Monte Carlo, risk-based surveillance systems, sample size calculator, exploratory analysis for space-temporal epidemiology and automatic abstract and full-text screening using machine learning. • Data collection framework: a data collection framework (DCF) is currently in place, responsible for collecting, validating and storing scientific raw data from Member States in relation to contaminants, pesticides, biological hazards and consumption data, as well as data regarding antimicrobial resistance, microbiological contaminants and food-born outbreak data in the contexts of zoonoses. The system provides support for all reference terminologies used in EFSA, such as the list of substances (used also for dossier management) and FoodEx2 (international standard used for food coding). • Management data tools: Currently scientific data are accessed, managed and analysed through a combination of systems, including ArcGIS (Geospatial information management system), MicroStrategy (Reporting tool) and SAS (Analytics tool). • Computing environment: a personal computing environment that runs remotely on the cloud exists and is accessible from any device for a limited number of scientific users that need access to the Scientific Data Warehouse applications. • Data protection: EFSA can only use cloud-based tools if the cloud provider agrees to sign a contract committing that they will host all the data within the European Union. Some feedback on Future Plans for EFSA has been identified as follows: • DATA 2.0 project: special importance is given to a project that EFSA is working on, named DATA 2.0. The project includes a tool which will be used to improve automation by converting data into standard formats that are widely accepted. • Data governance framework: a Data Governance Framework is being considered as a future project to support EFSA in terms of data accuracy, completeness, and consistency. 3.2.2.2. Target State requirements In this sub-section, a list of technical requirements and skills are provided with the aim of implementing the AI solutions identified during the data collection phase of the project. “Technical Requirements” refers to different terms, including infrastructure and tools needed for the development and training of AI tools, while “Skills” refers to terms such as knowledge of operating systems, software engineering and development. Technical Requirements • Data Infrastructure for Big Data and AI: The first technical requirement for EFSA's target state is a Data Infrastructure which can facilitate end-to-end Big Data and AI analysis and development. While many AI solutions are available in a Software-as-a-Service (SaaS) delivery model, and are thus easily accessible by any users or organisations, it is evident that for EFSA to be able to adopt solutions that fit the organisation’s requirements satisfactorily, and for moving into research activities, where data scientists need to develop or adapt a given algorithm, the expansion of EFSA’s current infrastructure to one that can address the new needs around Big Data and AI development is a pre-requisite. www.efsa.europa.eu/publications 32 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence • Programming Software: Common programming software for data preparation and data science includes SQL, R, Python, C# and Scala. • Subject-specific ontologies: To customise commercially available AI tools, or to develop them from scratch utilising Natural Language Processing (NLP) to address domain-specific applications, EFSA needs to adopt existing, or create from scratch, domain-specific ontologies (concepts and categories in a subject area or domain that shows their properties and the relations between them in the form of a table) Skills • Programming: In addition to knowledge of at least one of R, Python, C# and Scala, a good knowledge of fundamentals of Algorithms, Data Structures and Object-Oriented Languages is also needed, as well as the ability to code Quantitative and Statistical Analyses. • SQL – Structured Query Language: knowledge of SQL, Structured Query Language, is fundamental for extracting data from databases both in traditional data warehouses and big data technologies. • Data Visualisation: The ability to interpret the data visually (either within programming languages or via BI tools) is necessary in order for insights via analytics and AI analyses to be successfully conveyed. • Mathematical/Analytical Skills: Problem solving skills around complex mathematical/analytical background with an understanding of Big Data. • Knowledge of Natural Language Processing and Text Analytics: within the context of Evidence Management, EFSA’s target state requires data scientists with specialised skills and experience in Natural Language Processing (e.g., Named Entity Recognition, Text Summarisation, Topic Modelling, Semantic Search) and text mining. These skills are also needed for the development of domain-specific ontologies using state-ofthe-art text mining methodologies. • Combination of Analytical and Domain Expertise: For EFSA to be able to use AI/NLP across domains, it is important for data scientists to be able to acquire domain expertise so that the tools are programmed to answer specific scientific questions within a specialised domain. It is thus necessary for data science knowledge and skill to be dissipated within specific units rather than remain centralised. 3.2.2.3. data, require a solid Addressing challenges and blockers identified by the Gap Analysis In this sub-section, an analysis of possible solutions that might support EFSA in overcoming the challenges and blockers identified during the previous activities, is presented. Via the analysis of the current vs target state, we have recognised that for some of the identified gaps there is a good level of maturity within EFSA, however there is still room for improvement given that a series of tools, functionalities and organisational changes can provide further benefits. The identified gaps have provided the basis for the horizontal recommendations presented in section 4.1. Table 7 - Gaps and Recommendations per each dimension. Dimension Infrastructure Tools & Data Information Organisation Governance & Process Integration & & Gap While EFSA has adopted a cloud-based platform and uses multiple programming languages and a BI tool, the current infrastructure might not be sufficient for the use of Big Data and AI EFSA does not have any domain-specific ontologies While there are plans for a Data Governance framework, it is still not in place. Since AI has still not been widely adopted within EFSA, there are no guidelines in place to ensure its ethical and transparent use. Furthermore, for the same reasons, no ML model development management framework is in place. www.efsa.europa.eu/publications 33 Recommendation Implement a Data Infrastructure for Big Data and AI Create (or adopt existing) domainspecific ontologies Implement a Data Governance framework that considers AI Establish DevOps and adopt a Trustworthy AI framework EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Skills & Culture There is a plethora of data science skills within EFSA scattered across some teams, however there are still some skill gaps, especially related to big data technologies and NLP. Develop data science skills around NLP and Big Data and design behaviour change mechanisms The horizontal recommendations that can address these gaps are presented and elaborated in section 4.1. 3.3. Cooperation/collaboration opportunities This sub-section outlines and evaluates the cooperation/collaboration opportunities to be implemented with the identified potential partners. It first provides an identification of the phases of the end-to-end lifecycle, from strategy to operation and a list of possible collaboration/cooperation models to be adopted with multiple stakeholders. 3.3.1. AI Lifecycle Phases Third party collaboration/cooperation can happen in different phases of an AI adoption project, with different objectives and categories of stakeholders. For this reason, we introduced a conceptual framework that breaks down AI adoption into phases – we call it the “AI lifecycle” – and used this to better understand when collaboration/cooperation can happen, for what objectives, through what modalities and with which type of stakeholders. Our previous data collection activities have been functional to this exercise because they shed light on the different cooperation modalities available for EFSA, and on the composition of the stakeholder ecosystem. The AI lifecycle, instead, is part of the output of the previous project, “Joining forces at EU level on the implementation of Artificial Intelligence” (2019), and it includes all steps from strategy development to the monitoring of AI-based solutions. Specifically, the lifecycle can be broken down into 8 phases (as shown in Figure 13). The phases might occur in parallel or be re-iterated, especially if an agile development approach is followed. Their description is provided below. Figure 13 - AI lifecycle phases. • Strategy: consists of the design of EFSA’s AI strategy in line with the organisation's values, brand image, morals and ethics (including several tasks such as the definition of its scope across the organisation, ownership and accountability for its delivery and identification of impacted stakeholders’ groups), and guideline to the related implementation such as Portfolio Management, including market research for identifying and prioritising industry use cases and the definition of the desirable outcomes and risks for the use cases. Furthermore, this phase includes the Policy & Regulation area, including tasks such as the identification of relevant regulatory requirements and business policies relating to the use of AI, and definition of the expected businesses compliance requirements for the solution. • Delivery: consists of the processes to manage the use cases’ development in terms of Organisational Change, such as the development of plans and capabilities and definition of data science skills and business functional knowledge within business functions, to enable the adoption and support of AI solutions. Also, this phase includes the Programme Governance area, aimed at defining a clear program scope, roles, www.efsa.europa.eu/publications 34 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence responsibilities, desired business outcomes, quality standards and risks to be monitored and managed through several tools (e.g., quality assurance mechanisms). Finally, this phase foresees the adoption of an Agile Delivery Approach to encourage fail-fast learning and minimum viable product releases, and also the definition of a tailored AI development lifecycle. • Ecosystem: consists of processes for Vendor Management to evaluate different sourcing options (e.g., build vs buy options) and select the best vendor able to meet both functional and non-functional requirements, and to demonstrate the quality of the solutions. In addition, this phase includes the definition of a Technology Roadmap, involving the scouting for new technological options and their integration into the enterprise IT estate for production support. • Furthermore, it foresees the definition of Standards and Practices (e.g. for the ethical use of AI, and development of AI solutions through a consistent approach). A specific focus on vendor management within this phase, applied in the context of the AI-RM project, is provided in sub-section 3.3.3.1 • Data: consists of the definition of Data Requirements considering data type, quality, completeness, and volume, and development of Data Gathering process from desired data sources, as well as Data Preparation tasks to support AI models, leveraging on transformation functions, features selection, data labelling, and data sampling. In addition, all of these activities within this phase need to follow the adopted Data Governance framework. • Design: consists of the selection and development of a final AI Model (involving the also the definition of appropriate performance measures and prior use of cross-validation to estimate the model performance by means of training data), and its consequent Verification and Tuning to test the model in terms of robustness, and consistency and compare it against functional and non-functional requirements. • Also, this phase includes the definition of Controls Framework (e.g., COSO framework to measure the most prevalent risks), aligned with the ethics and governance guidelines and policies to ensure the quality of the solution. • Implementation: consists of the implementation of the AI solution managing Business Readiness through change management activities (i.e. making sure there are well-defined and shared use cases, and that the organisation is aware of the central role of data for AI adoption), Deployment of the solutions in the production environment following robust testing and errors management and Transition of the solution in the daily operations without any negative disruption, and making sure it is provided with the necessary hand over and/or training. • Operating: consists of the operating activities to guarantee Refinement & Improvement of the model’s performance through feedback collection mechanisms and review of the model’s outputs. Also, it includes the Resilience area, involving the definition of a business continuity plan and incident response procedures, and Support for the AI solution, involving the effective implementation of incident and problem management procedures, knowledge management procedures, and the change and release management procedure. • Monitoring: consists of the monitoring activities to ensure Compliance, involving tasks such as the implementation of checks on data protection, definition and use of a register with records of all processing activities, and provision of evidence of compliance with policies and regulations to third parties. Also, this phase includes the Cyber security area, including tasks such as the definition of effective methods to identity, authenticate and check access permissions of every data source/user/application/device connecting to data. Finally, this phase includes the Performance of the AI solution, consisting of the establishment of tools and mechanisms to evaluate the benefits, costs, and performance anomalies of the solution. Collaboration/cooperation with third parties can take place – and can be beneficial – in each one of these phases, having specific objectives that are related to the nature of the activity and the types of stakeholders that could be involved. The next paragraph adds another element to our reasoning, mapping the different www.efsa.europa.eu/publications 35 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence collaboration/cooperation models that an organisation like EFSA can establish. The AI lifecycle is also mentioned in the next section to contextualise the recommendations. 3.3.2. Map of potential cooperation/collaboration models As mentioned in the previous paragraph, this section provides an overview of the multiple collaboration/cooperation models that EFSA can establish with different partners for the potential adoption of AI solutions. According to the open innovation approach 16, multiple collaboration/cooperation opportunities can be established for the development and implementation of AI solutions within the evidence management process, with existing and potential stakeholders (from academia, industry and institutions etc.) composing the EFSA ecosystem (Figure 14). Figure 14 - State of Play of EFSA AI ecosystem of stakeholders. Some of these stakeholders have already built or are currently developing relevant solutions, capabilities or services that can be used for EFSA’s internal processes, or even have the necessary knowledge and expertise in AI in evidence management. However, other stakeholders are still in the research phase, or are designing an AI strategy or plan. The possible cooperation/collaboration models that EFSA could involve various scouting instruments and modalities to develop and deliver innovation. Table 8 provides a comprehensive overview of all cooperation models identified, together with some elements such as the main functions of the model under scope, and the stakeholders that can and may benefit from participating, together with a short description of the collaboration model. Table 8 - Comprehensive overview of cooperation models. Cooperation model Stakeholder Description Strategic alignment EU Agencies, EU Institutions, Refers to activities such as sharing ideas and lessons learned. Discussing common challenges with other public organisations can pave the way to the definition of a common vision and planning, and eventually to more concrete collaboration, avoiding 16 The open innovation approach is the use of purposive inflows and outflows of knowledge to accelerate internal innovation, while taking into consideration the opportunity to expand the markets for use of innovation (Eelko K.R.E.,Huizingh; Open innovation: State of the art and future perspectives; Technovation; 2011). www.efsa.europa.eu/publications 36 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Cooperation model Memorandum of Understanding Stakeholder Description International Organisations overlaps and waste of resources. Strategic alignment, therefore, is an important cooperation modality and it can take place through the organisation of conferences, workshops and ad-hoc meetings, such as those EFSA already does in the field of AI with the EU Agencies as part of the “Join forces” initiative. all Refers to an agreement between two or more parties that defines the framework of a negotiation, outlining the services to be exchanged and actions to be performed within a specific period of time. It expresses a convergence of will between the parties, indicating an intended common line of action, helping to explain any protocols for communication, information exchange, reporting, confidentiality issues, and modifications and conditions for terminating the agreement. MoUs are agreements in the form of legal documents. However, they are not fully binding in the way that contracts are, but they are stronger and more formal than a traditional gentleman's agreement. MODALITIES TO DEVELOP AND PROCURE INNOVATION: consisting in the launch of procurement procedures (of different kinds, depending on the TRL of the solution to be acquired) or the direct participation in projects or cooperation programmes for the development or the implementation Procurement Open (e.g. Industry players and academia) Refers to the process by which public authorities purchase work, goods and services. Usually results of the procured projects are owned by the contracting authority (unlike in the grant procedure, where is can be shared with the contractor). In this context, for instance, EFSA might launch a call for procurement in order to purchase off-the shelf solutions, procure related services for integration or maintenance and outsource internal activities. Procurement can also be published on behalf of two or more contracting authorities together (i.e. Joint procurement). Refers to a specific kind of procurement that includes: Innovation procurement Open, usually interesting for Industry and Academia e.g., startups, innovative SMEs, university spin offs i) the development of innovative solutions through the procurement of research and development (R&D) services (Pre-Commercial Procurement). In this case, the public buyer describes its need, prompting businesses and researchers to develop innovative products, services or processes to meet the need. ii) the procurement of innovative solutions that are not yet available or do not exist on the market, or the procurement of innovative solutions that do exist but are not yet widely available on the market (Public Procurement of Innovative Solutions). In this second case, Research & Development (R&D) of the solution is not needed. Different procedures could be applied to implement innovation procurement.17 Contest Open to public Refers to a recognition prize granted by a contracting authority to beneficiaries who can most effectively meet a defined challenge. Specifically, contests consist of acquiring a plan or design (e.g. of an AI prototype), to be selected by a jury, after being put out to competition. The aim is to stimulate innovation and come up with solutions to problems. The winner of a contest will receive cash, publicity coverage or promotion as a reward. Detailed Rules of contests are outlined in Art. 207 of the REGULATION (EU, Euratom) 2018/1046 OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL. 18 Grant Art 36 organisations Refers to a financial donation awarded by the contracting authority to the grant beneficiary. EFSA should consider awarding grants for projects and activities that contribute to EFSA’s mission, which consists in developing, implementing and 17 https://ec.europa.eu/growth/single-market/public-procurement/innovation-procurement_en Official Journal of the European Union; REGULATION (EU, Euratom) 2018/1046 OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of 18 July 2018 on the financial rules applicable to the general budget of the Union, 18 www.efsa.europa.eu/publications 37 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Cooperation model Stakeholder Description adopting AI within specific processes of the evidence management19. According to EFSA 20 there are several types of grants that the agency can launch: Specific grants, Tasking grants, Thematic grants, Fellowship grants, Partnering grants. Cooperation programme EU Agencies Refers to multi-agency programme that funds and manages various projects and joint actions, with a shared budget and governance. Within a Cooperation Programme, call for tenders and call for proposal can be launched with potential partners. Horizon Europe National food safety authorities, Academia Refers to participation to research & innovation projects carried out by consortia of organisations working together on specific areas. EFSA can be an interested party in those projects, therefore it can engage with the community, provide data for development and testing of solutions, and other activities, depending on the scope of the projects. SCOUTING INSTRUMENTS and IDEAS GENERATION: instruments which can be used to collect innovative ideas and solutions on AI Public consultation Internet crowdsourcing Hackathons Open to the public Refers, according to EFSA sources, to the creation of an effective exchange on a draft scientific output based on a decision of EFSA to seek comments from the public, namely the non-institutional stakeholders, which include academics, NGOs, industry and all other potentially interested and affected parties21. All relevant public and private stakeholders Refers to a central scouting activity that can give access to a wider set of expertise by engaging a broader spectrum of sources via the use of the internet and purpose-built platforms to elicit knowledge and services. Depending on the request, there are different types of crowdsourcing. A crowd contest is one of the most common types that an actor can use by seeking multiple qualified companies or individuals who will submit their own idea or solution (depending on the assigned task e.g., testing or data analysis). The winner gets funds for producing this solution or idea22. Open to the public Refers to the organisation of design sprint-like events together with other partners such as universities or industries, which can have the focus on design of AI solutions to be adopted for evidence management. Hackathons are scouting instruments that have proven to be an excellent medium to generate ideas for new tech products and services, and it can lead to the creation of innovative concepts, ideas, and prototypes while, at the same time, it stimulates the innovative ecosystem. According to the open innovation framework, several development levers could be taken into account to accelerate the modalities to engage the AI community and establish cooperation. These levers include the items reported in the table below (Table 9). Table 9 - Development Levers Development levers Description 19 https://ec.europa.eu/international-partnerships/grants_en https://www.efsa.europa.eu/en/discover/infographics/yes-we-grant 21Https://www.efsa.europa.eu/sites/default/files/event/documentset/5_public_consultation_final_document_110609.pdf 22 EFSA https://www.efsa.europa.eu/en/press/news/190614 20 www.efsa.europa.eu/publications 38 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Digital innovation hubs (DIHs) DIHs are at the core of European digital policy as they function as one-stop shops that help companies become more competitive with regard to their business/production processes, products or services using digital technologies, by providing access to technical expertise and experimentation, so that companies can “test-beforeinvesting”. They also provide innovation services, such as financing advice, training and skills development that are needed for a successful digital transformation. DIHs also support companies and public sector organisations in the use of digital technology to improve the sustainability of their processes and products. A DIH as a multi-partner cooperation can provide strong linkages with different actors inside and outside of its region promoting the implementation of multiple digital opportunities in the industry and the public sector. A partnership with a DIH focused on AI could help EFSA to engage the AI community and foster the uptake of digital technologies, including AI solutions. Technology transfer office Refers to an alternative approach to promote innovative solutions in line with the EFSA needs regarding the development and implementation of AI solutions within the evidence management process23. Establishing a partnership with a technology transfer office will promote the translation of recent technology advances into marketable solutions by maintaining close and long-term relationships with the results creators. Incubators and accelerators Refers to an organisation that helps start-up companies and individual entrepreneurs to develop their businesses by providing a full-scale range of services, starting with management training and office space and ending with venture capital financing. Incubators help many promising business ideas to be developed and take the first steps to transforming them into marketable and scalable ideas). A partnership with an incubator is strategic in nature as it will be a great source of innovative solutions that will bring new knowledge, skills and experience to the surface, which are necessary tools for the implementation of EFSA's aspirations Once the multiple cooperation/collaboration models and the related stakeholder category are defined (as mapped in Table 9), they can be associated to the different phases of the end-to-end AI lifecycle (described in section 3.1.1) based on the needs and added value for both EFSA and possible third parties in cooperation/collaboration, as shown in Figure 15 below. Figure 15 - AI lifecycle and collaboration models. Aaboen, Lise & Holgersson, Marcus. (2016). "Technology Transfer Offices, Incubators, and Intellectual Property Management". Academy of Management Proceedings. 2016. 23 www.efsa.europa.eu/publications 39 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence This mapping is further elaborated and discussed in section 5.2 “Engagement Plan”, where EFSA needs and the reason why (main motives and incentives) a specific stakeholder category should be willing to collaborate with EFSA through a specific model within a particular AI lifecycle phase are also defined. 3.3.2.1. Vendor selection and procurement management One of the key phases where collaboration with third parties is established is the “ecosystem” phase of the AI lifecycle, where, in the vendor management activities, an organisation defines - based on its needs - the best approach to procure or develop AI solutions to be integrated within its IT ecosystem. Being a crucial part, this paragraph focuses on this procedure and its possible outcomes, in order to allow EFSA to choose the best vendor selection procedure, and pave the way for the vertical recommendations (see section 4.2.). The vendor selection procedure for AI solutions is made up of the following steps: 1. Definition of the need - Definition of the challenge or need in terms of scope, objectives (benefits, performance and expected impact) and, if possible, requirements (functional and non-functional). In other words, the need identifies the use case to be improved thanks to the adoption of AI technology. 2. Market scouting - Research, based on the identified need, of the market solutions that could be ready to be purchased and integrated within the system to solve the challenge or need. The result of this step can be that either there are many solutions (≥ 3 solutions, for instance), or that there are none/very few. 3. Coverage assessment - If there are many available solutions, an assessment of the coverage of the solutions against the specific requirements should be carried out. This aims to understand whether they can cover fully (>90%), partially (90% ≥ x ≥ 60%) or insufficiently (<60%) the requirements defined at step 1. 4. Scenario selection - Based on the outcome of the previous steps, there are 3 options for the selection of the best vendor scenarios: • Scenario 1: If a high number (at least 3) of available solutions with a high coverage (higher than 90%) have been identified, the AI market is mature enough for the organisation to purchase a good off-the-shelf solution, without incurring in additional R&D costs. Therefore, a procurement procedure would be the most suitable model of collaboration, with the aim to purchase solutions that are ready to be tested and integrated into the production environment with little customisation. Benefits Launching a procurement procedure to buy off-the-shelf solutions is usually the most efficient – in terms of costs and time – and effective way for an organisation to adopt AI. Although even in this case the integration process could require some customisation activities, the project would not have high risks because the solutions already cover very well the needs. In addition, procurement would allow internal resources to focus on core organisational activity. Challenges Three activities are key to the launch of an effective procurement procedure for an AI solution to be purchased off-the-shelf: first, market scouting, which has to reflect real market potential and be able to assess with a high degree of certainty whether there are potentially fitting solutions to be bought; second, the stakeholder engagement activities around the opening of the procedure have to ensure that a good number of proposals are collected, in order to allow the organisation to choose the best one in terms of quality and costs; third, the tender itself has to be clear, and allow the potential contractors to well understand the needs, constraints and desired impact. For this reason, particular attention must be paid to these activities, which should be allocated to internal staff with the right skills, experience and understanding of the domain of application. In addition to the challenges related to the procurement itself, buying an offthe-shelf solution could entail the risk of a low degree of solution/vendor flexibility in case of interim changes of need (scope, requirements), although this can be mitigated by adopting an agile project management approach. www.efsa.europa.eu/publications 40 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence • Scenario 2: If from the market scouting and the coverage assessment at least 3 commercialised solutions with a medium level of coverage (between 90% and 60%) are identified, it would be risky to try to buy an off-the-shelf solution, as the procurement procedure may not be successful. With this in mind, the organisation could launch a procurement procedure including a PoC (Proof of Concept)/Pilot phase to demonstrate the quality and feasibility of the solutions before integration. Another option could be the launch of a PPI (Public Procurement of Innovative Solutions) , a particular type of innovation procurement, to purchase AI solutions that were not commercialised before or that still need some developments, or existing solutions that need customisation phases to match perfectly the need, including the PoC/Pilot phase as well. Benefits Innovation Procurement can be a flexible way to develop and implement highly innovative solutions for solving new problems, and to meet needs that are not fully covered by existing solutions in the market. A procurement that includes a PoC in the foreseen tasks has a better probability of achieving the objectives. Challenges Launching an innovation procurement might result in being a more time-consuming and expensive process in comparison to standard procurement, because of the need to perform a Proof of Concept/Pilot phase to demonstrate the solution's performance and feasibility. • Scenario 3: If few solutions (less than 3) or no solutions have been identified by the market scouting, or if the solutions identified cover insufficiently the needs (less than 60%), it will not be possible to buy any solutions from the market, and so a new solution, tailored on the needs, must be developed from scratch. If the organisation cannot/does not want to develop the algorithms internally, it can launch a grant to finance R&D activities aimed at developing and implementing AI-based innovative solutions that respond to needs and resolve problems that are not addressed by the solutions available on the market. In addition, EFSA may launch a PCP (Pre-Commercial procurement) to buy research and development (R&D) services from several competing suppliers, including the design, prototyping, original development and validation/testing of a limited set of first products. In this regard, it should be noted that a PCP does not include services related to the ‘commercialisation’ of the solution, intended as the full deployment/implementation and diffusion of a large scale of end-products within the organisation. Hence, following the PCP, EFSA could consider opening a PPI for the production and implementation of the solution (in larger quantities) to be adopted widely by EFSA 24. Finally, EFSA might consider opening a Contest procedure in order to develop, test and validate the development of an innovative solution. These scenarios would also involve a request to vendors to produce a PoC / Pilot Project, and the final integration of this within the selected process to be automated. Benefits In general, outsourcing development activities with open competitions ensures the identification of highly skilled teams at a convenient cost. With regard to EFSA grants, given that only organisations included in the list of Competent Organisations as designated by a Member State can participate, we can assume that they are aligned with EFSA's mission. In addition, opening a grant would also reinforce synergies among different stakeholder participating under Article 3625. Launching a PCP would facilitate access to new innovative players, such as start-ups and SMEs, and to the public procurement market, sharing the risks and benefits of designing, prototyping, and testing new products and services and creating the optimum conditions for wider commercialisation at a later stage. It 24 It should be noted that EFSA may engage the same vendor of the PPI by defining specific criteria within the PCP call, or by only opening the PCP call to selected vendors. 25https://efsa.force.com/competentorganisations/s/recordlist/CompetentOrganisation__c/00B1v000009LqfIEAS www.efsa.europa.eu/publications 41 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence is important to note that, in a PCP, suppliers usually retain IPR ownership rights, while procurers keep some use and licensing rights. Finally, opening a Contest would be an opportunity to reinforce synergies among the EFSA community. In addition, it would enhance the reputation of EFSA about its mission and strategy, including key value such as openness and innovation. Challenges In general, considering the low level of readiness of the solutions, risks related to the loss of control on the quality and performance of the activities might emerge with a grant, PCP or Contest, together with a higher investment in development activities and tests, which could require more time and effort than expected. At the same time, grants would restrict the call to a limited number of stakeholders, considering, in fact, only organisations within Art. 36, entailing the risk of not finding the right proposals of the solution to be adopted by EFSA. Indeed, the estimation for consequently opening both the procedures would require at least a year and half. The entire process is represented graphically in Figure 16 below. Figure 16 - Vendor selection and procurement management process. In the context of this project, we have applied this approach to the 10 prioritised AI use cases of the evidence management process, as output of the vertical recommendation, relying on the market scouting carried out during the data collection activities (section 4.2). However, EFSA will be able to apply this methodological framework to select the best vendor scenario in all use cases, including those that were not prioritised (see long list of use cases in sub-section 2.4.2). 3.3.3. Potential grounds for collaboration As mentioned in the previous sections, collaboration opportunities throughout the AI lifecycle are multiple. In order to better assess these opportunities and reflect on who could be the most relevant partners with the most promising solutions, two exercises are presented in the following sections: firstly, the prioritisation of institutional partners based on a set of criteria and, secondly, a presentation of an overview of solutions that are currently commercialised, under development or in research phase, managed by academic, industrial and institutional stakeholders, for each prioritised use case. Going into more detail, the prioritisation of partnerships (sub-section 3.3.3.1) aims to assess, based on the information collected in the data collection activities, the potential grounds for collaboration in the context of the www.efsa.europa.eu/publications 42 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence adoption of AI in evidence management. We have performed this exercise with the institutional type of stakeholders, as collaboration with them does not have to go through vendor management, but can be oriented and negotiated. The approach can be applied to any institutions, and we have considered those that, according to data collection, have an interest in adopting AI in the evidence management process. Article 36 organisations are not present because the literature review and survey activities did not gather relevant/sufficient information. On the other hand, the overview of solutions developed by industrial, academic and industrial players – clustered per specific prioritised use case (sub-section 3.3.3.2) - aims to provide a comprehensive overview of stakeholders that could collaborate with EFSA through different modalities such as procurement, innovation procurement and grant. 3.3.3.1. Assessment of collaboration opportunities with other Institutions European Institutions, National Authorities and International Organisations, mainly engaged through the interviews, have been prioritised based on two criteria with the same weight: 1) Common needs/strategy: This criterion takes into account whether the organisations under scope have developed a strategy/roadmap that focuses on AI, whether they have the same needs and interest in relation to the development of AI in Evidence Management, or whether they have the same processes in EM with prioritised areas that require intervention. 2) Potential for collaboration: This criterion evaluates the ease of collaboration with the organisations, taking into consideration many factors such as whether the organisation under scope has already collaborated with EFSA, or if it is currently collaborating, and whether the organisation is an EU institution or an Article 36 organisation. Both criteria can have a 1-to-3 score (low, medium, high). Table 10 presents the prioritisation of institutional stakeholders on the basis of the criteria previously mentioned. In addition, feedback provided during the interviews and the workshop has been taken into consideration. Potential collaboration models for each organisation are also provided. For more detailed information concerning the description of the initiatives and solutions, see Table 3 in sub-section 3.1.1. Table 10 - Institutional organisation prioritisation. Institution CDT EUIPO EMA 50% 50% Common Strategy/ Needs Potential for collaboration 3 3 3 www.efsa.europa.eu/publications 3 3 3 Ranking Additional information 3 Based on the evidence presented in section 1.1.1, there are grounds for collaboration for a specific use case (automatic text summarisation), which could be carried out through joint procurement or a cooperation programme. Feedback provided during the interviews and the workshop was positive. 3 Based on the evidence presented in section 1.1.1, there are grounds for collaboration for some use cases (automatic text summarisation and automated table extraction). EFSA and EUIPO could cooperate through joint procurement or a cooperation programme. Feedback provided during the interviews and the workshop was positive. 3 Based on the evidence presented in section 1.1.1, there are grounds for collaboration due to applications of AI for text data structuring. EFSA and EUIPO could cooperate through joint procurement or a cooperation programme. Feedback provided during the interviews and the workshop was positive. 43 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Institution 50% 50% Common Strategy/ Needs Potential for collaboration Ranking Additional information Based on the evidence presented in section 1.1.1, on past collaborations between the two organisations, and on the fact they both deal with similar fields, there is room for collaboration, which could take place through a grant. Feedback provided during the interviews and the workshop was positive. VKM 3 3 3 ENISA 3 2 2.5 Based on the evidence presented in section 1.1.1, ENISA is quite interested in developments concerning AI and ML, but a specific use case of common interest has not been found. Further alignment would be needed to work together in a joint AI project. 2 Based on the evidence presented in section 1.1.1, EPA has worked on a tool that presents functionalities that resemble the use case abstract screening. In addition, EPA and EFSA work in very linked fields (environment and food safety). The challenge could concern the means to cooperate with a non-European Agency in concrete joint projects, therefore further strategic alignment would be needed. 2 Based on the evidence presented in section 1.1.1, FRA seems to focus more on the impact of AI for fundamental rights than to the application to specific use cases within the organisation's processes. Strategic alignment is anyway important to allow EFSA to advance with AI adoption taking into account FRA's guidelines. Feedback provided during the interviews and the workshop was positive. 2 Based on the evidence presented in section 1.1.1, the JRC has different research lines on AI, but seems not to be working on AI application within its processes. Strategic alignment with this institution is anyway important to eventually find grounds for cooperation in the future. Feedback provided during the interviews and the workshop was positive. 2 Based on the evidence presented in section 1.1.1, AMNH has a high degree of alignment concerning AI in EM (a representative from the organisation presented a tool that automates SLR and performs keywords identification). On the other hand, being an organisation that operates in the US, working on common projects would be more difficult. 1.5 Based on the evidence presented in section 1.1.1, OECD has different research lines on AI, but it has not developed any roadmap or solution dealing with EM. Strategic alignment in the field of AI could be established, to better evaluate if grounds for more specific collaboration arise in the future. 1.5 Based on the evidence presented in section 1.1.1, FAO seems to focus more on the use of AI in agriculture (for projects in the field) rather than in internal processes. Strategic alignment could be established, to better evaluate if grounds for more specific collaboration arise in the future. EPA FRA JRC American Museum of Natural History OECD FAO 3 1 1 3 1 1 1 3 3 1 2 2 In Figure 17 a prioritisation matrix with institutional stakeholders is presented, with organisations ranked according to the criteria previously presented. www.efsa.europa.eu/publications 44 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 17 - Prioritisation matrix of institutional stakeholders. 3.3.3.2. Collaboration opportunities per prioritised use case This section provides EFSA with a view on the relevant academic, industrial and institutional stakeholders with whom EFSA could collaborate through different methods and means for the adoption of AI technology in Evidence Management. The table below lists all identified solutions during the data collection activities (desk research, survey, interviews and the workshop) for each prioritised use case, together with their respective developer organisation’s name and technology readiness level based on the available information. Low readiness solutions refer to those solutions that are currently in a research phase, while medium readiness solutions refer to those that are currently under a development phase and, finally, high readiness refers to those solutions that are already commercialised, or ready to be used. Table 11 - Collaboration Opportunities per Prioritised Use Case. Prioritised use case Data Collection Terminology Assessment SLR Keywords identification Organisation Solutions offered Octopai Foundation for Research Technologies - Hellas Seven Past Nine Universite de Tours Uninova University of Science Technology Bannu Arab Academy for Science University of Manchester Scipilot SGS Digicomply Iris.ai Sciome EMA Risklick AG University of Edinburgh www.efsa.europa.eu/publications 45 Octopai and and Readiness level of solution H VisTA H Seven Past Nine Heterotoki Research Project Homogeneous Ontology Mapping Word Encoding Method HYPHEN Syras Search Engine Results Page Iris.ai Swift Review Internal tool Risklick AG Automated Systematic Search Deduplicator H H M M M L H H H H H H M EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Prioritised use case Organisation Solutions offered Agriculture and Agri-food Canada TU München SLR Relevant Screening Abstract Sciome Evidence Partners Seven Past Nine University of Manchester SLR Literature duplication De- Text Summarisation Rayyan AMNH VKM Eppi-Centre Pico Portal Sciome Evidence Partners Risklick AG Iris.ai ASReview Research Screener Robot Reviewer CDT EUIPO University of Edinburgh Agriculture and Agri-food Canada Allegheny College Macquarie University Rockfeller University SLR Literature Clustering Findings Gujarat Technological University Iris.ai Innovamol Consulting Eppi-Centre Scipilot SGS Digicomply Sciome Evidence Partners Risklick AG Crème Global LibrAlry Seven Past Nine University of Edinburgh Agriculture and Agri-food Canada www.efsa.europa.eu/publications 46 Exploring Identifiers of Research Articles Related to Food and Disease Using Artificial Intelligence Integrative Approach for Automated Literature Reviews Swift Review DistillerSR Seven Past Nine Systematic review workload reduction through certaintybased screening Rayyan Colandr Internal tool Eppi-Reviewer Pico Portal Swift Review DistillerSR Risklick AG Iris.ai ASReview Research Screener Robot Reviewer Internal tool Internal tool RegEX Exploring Identifiers of Research Articles Related to Food and Disease Using Artificial Intelligence Automated text summarisation solution Automated text summarisation solution Automated text summarisation solution Encoder-Decoder Iris.ai Innovamol Eppi-Reviewer Syras Search Engine Results Page Swift Review DistillerSR Risklick AG Crème Global LibrAlry Seven Past Nine RegEX Exploring Identifiers of Research Articles Related to Readiness level of solution M L H H H L H H H H H H H H H H H H H H M M M M M M H H H H H H H H H H M M M EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Prioritised use case Organisation Solutions offered University of Piraeus Universität Siegen University of California University of Manchester University of Manchester Quality Improvement of the Scientific Output Expert Pool Identification Expert Selection GINGER GradePro ProWritingAID National Taiwan University University Politehnica of Bucharest Harbin Institute of Technology Changchun University Pomato Department of Scientific Research, Visibilia SP Ramrao Adik Institue of Technology University of Electronic Science and Technology of China Textkernel MosaickTrack Springer Nature Ideal Pomato Leoforce IMT Atlantique Department of Scientific Research, Visibilia SP Ramrao Adik Institue of Technology University of Electronic Science and Technology of China Agriculture and Agrifood Canada Comparable Detection Appraisals Wageningen University & Research Information Technology University of San Francisco Food and Disease Using Artificial Intelligence Classification of dermatological images using advanced clustering techniques Clustering approach Machine learning to advance synthesis RobotAnalyst Systematic review workload reduction through certaintybased screening GingerSoftware GradePro ProWritingAID DISA ReadME Readiness level of solution M M M M M H H H M M Write-righter Writing Evaluation Model Pomato M M H Job Recommendation M Job Recommender System M Machine Learned ResumeJob Matching Textkernel MosaickTrack Reviewer Finder AI for Recruiting Pomato Arya Textkernel M H H H H H H M Job Recommendations M Job Recommender System M Machine Learned ResumeJob Matching Exploring Identifiers of Research Articles Related to Food and Disease Using Artificial Intelligence Automatic Systematic Literature Review Critical Appraisal Tools M M M M Concerning the prioritised use cases, Expert pool identification and Comparable Appraisals Detection, few developed and commercialised solutions were identified. These were mainly developed by small software companies and academic organisations. On the other hand, several solutions are being developed by industry and academic organisations, whereas no solutions provided by institutional organisations can be associated to these use cases. www.efsa.europa.eu/publications 47 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Taking into consideration the information provided above, and on the basis of the methodology provided in subsection 3.3.2.1, due to the low number of commercialised solutions and the high number of medium and low readiness solutions for these uses cases, EFSA could launch a grant or a PCP to finance R&D activities for the design and implementation of AI solutions that respond to its needs, thus resolving problems that are not addressed by solutions available on the market. On the other hand, for the use cases SLR Keywords identification, SLR Literature De-duplication, Text Summarisation, SLR Literature Findings Clustering, Quality Improvement of the Scientific Output and Expert Selection, Data Collection Terminology Assessment, SLR Relevant Abstract Screening, several developed and commercialised solutions were identified, mainly developed by small software companies and academic organisations. In addition, for these use cases, several solutions still under development or in their research phase have emerged, mainly managed by academic organisations. Taking into consideration the large number of commercialised solutions for these use cases, and following the methodology provided in sub-section 3.3.2.1, EFSA could collaborate with these organisations through different cooperation/models – procurement, innovation procurement or grants, based on the use case coverage level (i.e., how much the solution is able to fulfil the needs and requirements associated to the respective evidence management step). The details and outputs of the use case coverage assessment are presented in section 4.2. www.efsa.europa.eu/publications 48 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence 4. Recommendations for projects This chapter outlines and prioritises a list of recommendations to develop a harmonised approach on the implementation of AI methods in the evidence management phase of risk assessment. Two types of recommendations have been developed: • Horizontal recommendations, presented in section 4.1, cover foundational elements enabling a systematic and smooth adoption of customised AI-based solutions within the evidence management process, building the capabilities to adopt and implement AI within both the risk management process and within the organisation as a whole. They address several topics such as the infrastructure, governance and skills, that need to be in place to unlock the adoption of AI in specific use cases (vertical recommendations) and achieve EFSA’s vision and Strategy 2027 of adopting human-centric AI within the evidence management phase of risk assessment. • Vertical recommendations, presented in section 4.2, propose the optimal vendor management scenario and sourcing option (according to the approach described in sub-section 3.3.3.1) to acquire or develop AI solutions for each of the prioritised use cases of evidence management. These recommendations are designed as single projects that EFSA can decide to carry out – either in-house or through outsourcing– to improve the evidence management process, achieve its 2027 strategic objectives and prepare the organisation to embed AI technologies within its processes. In order to facilitate decision making, we have carried out a time and budget estimation and a prioritisation exercise in section 4.3, aimed at understanding which actions (both horizontal and vertical) can be performed in parallel, and which are instead dependent on the completion of other actions for their proper initiation. Figure 18 - Vertical and Horizontal recommendations link. These recommendations can be associated to different phases of the AI lifecycle (sub-section 3.3.2), as shown in Figure 19 below. www.efsa.europa.eu/publications 49 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 19 - Link of Recommendations with AI Lifecycle Phases. 4.1. Horizontal Recommendations To achieve its vision of increasing the accessibility and the breadth of the body of evidence, as well as to apply human-centric AI in close coexistence with existing human expertise by 2027, EFSA should bolster its preparedness by taking steps towards better data management. In particular, data management should be appropriate, not only for the types of data currently processed, but also for those data types planned to be processed in the future. Adopting a fit-for-purpose data infrastructure, tools enabling Big Data and AI development, and a data governance framework that addresses AI, while taking into consideration people, processes, and policies, are foundational elements towards a more robust structure, aligned with EFSA’s organisational goals. Horizontal recommendations include transversal considerations on which EFSA can build its readiness for an effective AI adoption, not only in evidence management, but also across the organisation’s units and teams. While EFSA may decide to adopt some ready-made solutions to address specific needs and use cases (see Vertical Recommendations), a Design-to-Value approach is proposed, whereby quick wins are obtained in parallel with capability development, such that EFSA is in a position, in the future, to build customised AI tools that will serve its own needs in a more tailored way, ensuring maximum impact for the organisation. Based on the Gap Analysis presented in section 3.2, recommendations are presented to address different dimensions of AI capability development: Technology & Infrastructure, Data & Information, Organisation & Governance, Process & Integration and Culture & Talent (see figure 20 below). To this end, the following initiatives and projects to support EFSA in achieving its vision are recommended: Figure 20 - Dimensions and horizontal recommendations. www.efsa.europa.eu/publications 50 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Recommendation 1: Develop or adopt existing ontologies for domains of relevance for EFSA Description and scope One of the key gaps that has emerged from the series of data collection activities conducted is the need to develop (or adopt) tailor-made ontologies for the specific domains that are related to EFSA's work. Several of the use cases examined rely on the use of Natural Language Processing (NLP) and text mining, and require the use of knowledge sources for the purpose of understanding, extraction and retrieval from unstructured text. Simpler tasks, such as named entity recognition (NER), which can be the basis of Automatic Text Summarisation, and SLR Literature Findings Clustering, can be simply based on terminologies. However, more complex tasks that rely on the identification of concepts (such as SLR Abstract Screening and Comparable Appraisals Detection), require knowledge bases in the form of structured ontologies that encode multiple lexical representations in natural language. An ontology formally represents knowledge as a set of concepts within a domain, as well as the relationships between these concepts (e.g., concepts that mean the same thing but may be expressed differently). In essence, an ontology includes a vocabulary of terms and specifications of their meaning. This includes definitions and an indication of how concepts are inter-related, which collectively impose a structure on the domain and constrain the possible interpretations of terms26. For EFSA, the development of accurate and complete domain-specific ontologies is an essential element to support the implementation of AI technologies to automate and improve several steps within the evidence management process. Indeed, the majority of AI-based tools/methods to be potentially adopted (even some of those commercially available) require customisation, as they need to be trained on a comprehensive subjectspecific ontology. This will improve the effectiveness of both supervised and unsupervised AI tools. Key considerations for ontologies development A key consideration for NLP is that an ontology needs to be complete with respect to the entities represented, as well as the relationships between them and natural-language synonyms27 For example, to automatically retrieve documents that discuss genotoxic carcinogens, an NLP system requires an existing ontology that has an identifier for genotoxic carcinogens that is correctly linked to the natural language terms (e.g., ‘genotoxic carcinogens’, 26 Gruber,T. (2008), Ontology, Encyclopedia of Database Systems Kaihong Liu, William R. Hogan, Rebecca S. Crowley (2011),Natural Language Processing methods and systems for biomedical ontology learning, Journal of Biomedical Informatics, Volume 44, Issue 1, pp 163-179. 27 www.efsa.europa.eu/publications 51 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence ‘mutagenic’, ‘carcinogenic’, etc.), and also has relationships with additional identifiers for other entities, such as genotoxic carcinogens interact with DNA. As a consequence, when an ontology lacks a representation of an entity, a particular term associated with the entity, or some of its particular relations, the quality of an NLP system based solely on such ontology will be negatively affected. Therefore, the lack of any representation of an entity inhibits the detection of such entity. Equally, in the event that a given document uses a synonym to refer to a concept, the lack of a synonym present in the ontology would prevent the recognition of the entity. Lastly, the lack of an entity’s relationship might prevent and identification of the right answers to specific request as, for example, How do genotoxic carcinogens cause cancer? In more detail, an Ontology can be defined as a set of Individuals, Concepts (Classes), Attributes, Relations, Instances and Axioms: • Individuals: represent the basic, "ground level" components of an ontology. The individuals in an ontology may include concrete objects such as people, animals, tables, automobiles, molecules, and planets, as well as abstract individuals including numbers and words; • Concepts (Classes): represent a set or class of entities, or `things/objects’ within a domain (e.g., Protein is a concept within the domain of molecular biology), which can be of two types: • o Primitive concepts: are those which only have necessary conditions (in terms of their properties) for membership in a given class28; o Defined concepts: are those for which the use of a description is needed to be classified under a class29. Attributes: include aspects, properties, features, characteristics, or parameters that objects (and classes) can have; • Relations: represent the interactions between concepts or concepts’ properties, and can be of two types: o Taxonomy: organising concepts into sub-concept tree structures. A common form of taxonomy is the specialisation relationship (e.g., Enzyme is a kind of Protein, which in turn is a kind of Macromolecule); o Associative Relationships: relating concepts across tree structures. A common form is the nominative relationship that describes the names of concepts (e.g., Protein has Accession Number, in the context of bioinformatics) • Instances: are describe as real-world examples represented by a concept. • Axioms: are considered as a set of rules used to limit the values within a class in a specific domain of application. The process of ontology development is still to a large extent manual. For example, identifiers, their synonyms and relationships are still added one by one manually. The investment in ontology development in other areas has been huge, with examples such as the Gene Ontology (GO), a Consortium funded by the National Human Genome Research Institute30 , and the National Center for Biomedical Ontology (NCBO)31, amongst others. 28 For example, a globular protein is a kind of protein with a hydrophobic core, so all globular proteins must have a hydrophobic core, but there could be other things that have a hydrophobic core that are not globular proteins. 29 For example, Eukaryotic cells are kinds of cells that have a nucleus. Not only does every eukaryotic cell have a nucleus, but every nucleus also containing a cell is eukaryotic. 30 National Institutes of Health. Research Portfolio Online Reporting Tools (RePORT), 2010. Available from: http://projectreporter.nih.gov/project_info_history.cfm?aid=7941562&icde=2611544. 31 BioInform. Stanford’s Mark Musen on the New National Center for Biomedical Ontology, 2005. Available from: http://www.genomeweb.com/informatics/stanford-s-mark-musen-new-national-center-biomedical-ontology. www.efsa.europa.eu/publications 52 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence It is worth noting that a significant set of resources is already available to provide a foundation for the development of ontologies of relevance to EFSA’s work such as: • The Open Biological and Biomedical Ontology (OBO) Foundry32, an open-source initiative that aims to develop a family of interoperable ontologies, which are both logically well-formed and scientifically accurate. Through this initiative, participants contribute to the development of an evolving set of principles including open use, collaborative development, non-overlapping and strictly-scoped content, and common syntax and relations, based on ontology models that work well, such as the Gene Ontology (GO). • The ECOTOXicology Knowledgebase (ECOTOX)33, a comprehensive, publicly available Knowledgebase providing single chemical environmental toxicity data on aquatic life, terrestrial plants and wildlife. • The OpenTox ontology34, an ontology of toxicological endpoints, used for the analysis and prediction of toxic events. • The Gene Ontology (GO)35 is the world’s largest source of information on the functions of genes. This knowledge is both human-readable and machine-readable, and is a foundation for computational analysis of large-scale molecular biology and genetics experiments in biomedical research. • Other research efforts and initiatives that have been focused on the creation and testing of ontologies for toxicologic data, by merging a wide range of public bio-ontologies relevant to toxicology36, including the QSAR toolbox, eTOX, Pistoia Alliance, ToxWiz, Virtual Liver, EU-ADR, BEL, ToxML, and Bioclipse37. In addition to the already available open-source ontologies that EFSA can take advantage of, Protégé38, a free, open-source ontology editor and knowledge management system developed by the Stanford Centre for Biomedical Informatics Research at the Stanford University School of Medicine that can be leveraged for user-friendly ontology development should be considered too. Protégé, in fact, allows the definition of classes, class hierarchy’s variables, variable-value restrictions, and the relationships between classes and the properties of these relationships. Overall, it is considered as one the most widely deployed modelling tools, and the leading ontology editor. Protégé allows the definition of classes, class hierarchy’s variables, variable-value restrictions, and the relationships between classes and the properties of these relationships. Recommendation Summary Objectives - The recommendation suggests the development and adoption of well-structured ontologies within EFSA in order to improve the performance of AI tools to be acquired/developed and used. The adoption of ontology ensures that EFSA can make use of domain knowledge and scope, rather than the general operational knowledge that is currently being used. Challenges - 32 33 34 35 Given that building a solid and functional ontology depends on the complexity of the problem and the knowledge of the respective domain, experts within the organisation or externally outsourced must be consulted and involved when developing the ontology https://obofoundry.org/ https://cfpub.epa.gov/ecotox/ See http://old.opentox.org/dev/Ontology http://geneontology.org/ 36 Wang RL, Edwards S, Ives C. Ontology-based semantic mapping of chemical toxicities. Toxicology. 2019 Jan 15;412:89-100. doi: 10.1016/j.tox.2018.11.005. Epub 2018 Nov 20. PMID: 30468866; PMCID: PMC6774258. 37Hardy B, Apic G, Carthew P, Clark D, Cook D, Dix I, Escher S, Hastings J, Heard DJ, Jeliazkova N, Judson P, Matis-Mitchell S, Mitic D, Myatt G, Shah I, Spjuth O, Tcheremenskaia O, Toldo L, Watson D, White A, Yang C. 38 https://protege.stanford.edu/ www.efsa.europa.eu/publications 53 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence - Since the ontology development is a challenging and time-intensive task, it should be ensured that domain experts (potential ontology authors) are aligned about the usefulness and benefits of building and using an ontology, and specifically, how it can positively affect the performance of the AI tools. Expected Impact - An ontology can streamline the AI training process by automatically tagging, categorising and linking information to improve performance. An ontology can enable a smoother data integration and alignment if trained on different data sources. The Development and adoption of ontologies are foundational elements for building the appropriate infrastructure and tools to achieve improved data extraction, integration and standardisation to support the use of NAMs in chemical risk assessment. The following table provides the SWOT analysis for Recommendation 1 SWOT Analysis Strengths - Opportunities A number of ontologies are already available in open-source databases that can be reused by EFSA to perform pilots and assess the improvement of the performance of AI tools. Some of these ontologies have been built via multi-partner consortia and are fairly complete for specific domains. - Weaknesses - An ontology can be a decisive factor in improving the algorithm of the AI tool, by providing a dynamic knowledge-based that would be more preferable than static frameworks. Threats No advanced expertise in the ontology engineering field currently exists within EFSA, which is responsible for setting the activities that concern the ontology development process and lifecycle. - - The purpose of the ontology could be misunderstood, and it could be merely conceived as a tool to structure data, like taxonomy. As a result, this may affect the quality of the data within the ontology, leading to inaccurate training of the AI tool. It might be difficult to develop ontologies to cover the wide spectrum of EFSA’s domains of expertise. There is a need for maintenance and systematic updating of the ontology, which is recommended every two years, depending on how quickly scientific areas evolve (i.e., how fast new terms and concepts are developed) Recommendation 2: Implement a Data Infrastructure for Big Data Analytics & AI development Description and scope In order to develop and adopt an efficient Data Governance framework, EFSA should consider building a consistent Data Infrastructure. Designing and implementing a fit-for-purpose and scalable data infrastructure with the right technology and tools should lay the grounds towards a more effective data accessibility, integration and analysis, which might facilitate the use of AI in the evidence management phase of the risk assessment process and, at the same time, would also support EFSA towards a holistic adoption of AI and Big Data analytics. In this sense, this recommendation does not exclusively focus on the scope of evidence management, but it addresses EFSA as a whole organisation, considering the inputs collected from other EFSA units and the needs identified in other roadmaps, such as NAMs and PERA. www.efsa.europa.eu/publications 54 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Going into more details of this recommendation, EFSA needs to implement a Data Infrastructure to enable Big Data Analytics and AI for the following reasons: • The volume of data that EFSA will need to store and process in the future is likely to be too large for a traditional database. • A large part of the data will consist of unstructured data (e.g., textual data) that need to be accurately transformed for some key activities within the organisation, such as data analysis and reporting. • Real-time data collection and processing may be considered in EFSA’s future, especially in the context of an evolving and more dynamic risk assessment process, and in relation to the future possibility to use external unstructured data (e.g., from social media and news sites). There are various data architecture frameworks of different complexities that can be used. Here, we present the simplest form of data architecture, which contains three fundamental inter-connected layers, each characterised by specific elements and procedures: • Storage Layer: in this layer, data is obtained and ingested from multiple internal and external sources and then converted into a suitable format to enable a correct analysis. Depending on the type of data, the appropriate storage model must be used. Furthermore, it is worth mentioning that all of the data coming from the different data sources and available for analysis are appropriately selected, before moving into this layer. • Process Layer: in this layer, data is read and analysed from the storage layer to produce meaningful information. Two types of data processing can be distinguished: Batch processing, where processing happens to a set of data that has already been stored over a period of time in order to obtain detailed analytics and insights, and Stream processing, where processing happens while data is stored (continuing flow of data), to quickly generate real time insights through analytics tools. • Visualisation Layer: in this layer, comprehensive reports and visualisations to support decision making are created, based on the output of the process layer. Indicative Data Infrastructure Blueprint A blueprint of a data infrastructure is set out below (Figure 21), which is well suited to EFSA's needs in the area of evidence management within the risk assessment process. The blueprint can also be used to fulfil EFSA's wider needs around the domain of Big Data and AI across multiple units within the organisation. Specifically, it includes the different logical components that fit into a big data infrastructure. It is worth noting that EFSA's final layout may not contain every item in this diagram, on the grounds of specific organisation's short-term and long-term goals, anticipated types and volumes of data, and the required types of data analysis. Figure 21 - Indicative Data Infrastructure for EFSA www.efsa.europa.eu/publications 55 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Furthermore, depending on the vendor that EFSA may choose for the adoption of an AI-based solution, different services of the infrastructure may be selected. In Figure 22 below, the indicative data infrastructure is represented and integrated with the Microsoft Azure cloud platform's components, which is EFSA's cloud platform currently in use. Figure 22 - Indicative Data Infrastructure for EFSA based on the Azure cloud platform. Expanding EFSA’s current cloud platform solution with additional components to fulfil the increasing demand would lead to time and cost efficiencies. Below, we provide some additional details and key considerations about the different components of the Data Architecture that should be considered by EFSA: - Data Sources: an important topic of consideration for EFSA is the need to integrate multiple and possibly disparate internal and external data sources. The topic of data integration has been recognised as a pain point by multiple stakeholders within EFSA, including the NAMs project team. These data sources may consist of structured data (e.g., scientific measurements) or unstructured data (e.g., text from papers, other documentation or external unstructured data sources such as news and social media). - Storage Layer - Ingestion/Data Storage: it is important to ensure consistency in the data collection and process step through the selection of an advanced tool that can store high volumes of large files in various formats. This kind of storage is usually provided by a system/repository defined as a data lake. Possible options for implementing this storage include Azure Data Lake Store, or blob containers in Azure Storage. Most of the big data solutions involve various activities such as repeated data processing operations, encapsulated in workflows, which transform source data, move data between multiple sources and sinks, load the processed data into an analytical data store, or push the results straight to a report or dashboard. To automate these workflows, an orchestration technology such as Azure Data Factory can be used. www.efsa.europa.eu/publications 56 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence In this regard, the process of automating workflows to perform Extract Transform and Load (ETL) requires particular attention and consideration from EFSA. ETL is a data integration process that combines, cleans, transforms, and standardises data from multiple data sources and loads it into a data warehouse or other target system, in a way that addresses specific business intelligence needs. In particular, multiple EFSA stakeholders consulted in the previous phases of the project, highlighted the importance of data standardisation as a pre-requisite and foundation for data analytics and machine learning workstreams. Process Layer - Analytics/machine learning: it is recommended to select a big and advanced data solution that will process the data files using long-running batch jobs to filter, aggregate, and prepare the data for a proper analysis. Usually, these tasks involve reading source files, processing them, and writing the output to new files. Therefore, possible options include running U-SQL jobs in Azure Data Lake Analytics, using Hive, Pig, or custom Map/Reduce jobs in an HDInsight Hadoop cluster, or using Java, Scala, or Python programs in an HDInsight Spark cluster. In addition, Azure Machine Learning Studio can be used to build and deploy the machine learning models using two approaches: either Python scripts or with Hive tables on an HDInsight (Hadoop) cluster. - Concerning, natural language processing capabilities, these are offered either by Azure HDInsight, with Spark and Spark Mllib, or by Microsoft Cognitive Services. Regarding the latter, it contains pre-built NLP models that can be used on free text, whereas the former gives the option to train customised models against a large corpus of text data. Microsoft Cognitive Services would be a more appropriate choice for EFSA, since it provides more high-level NLP capabilities and, in addition, it can be customised to develop use cases like SLR (Systematic Literature Review) information clustering and text summarisation, without the need for significant coding and customisation. Visualisation Layer - Reporting: the goal of most of the big data solutions is to generate insights from the data stored, through accurate analysis and reporting. To empower users in analysing data, the architecture may include a data modelling layer, such as a multidimensional OLAP cube or, tabular data model in Azure Analysis Services. It might also support self-service Business Intelligence, using the modelling and visualisation technologies in Microsoft Power BI or Microsoft Excel. - To conclude, the following table presents a summary of recommendation 2 (Implement a Data Infrastructure to enable Big Data Analytics and AI), by outlining its objectives, challenges, expected impact, and a final swot analysis. Recommendation Summary Objectives - Design and adopt a Data Infrastructure that will fulfil EFSA’s current and future data analysis needs within the evidence management process and across the entire organisation. Effectively manage data and its flow throughout multiple infrastructural layers within EFSA. Build the capability to store and process big data from different sources, including unstructured data. Select a scalable and fit-for-purpose technological solution, such that EFSA’s changing needs can be met with minimal disruption. Challenges - Storing and processing big data from different sources, e.g., the collection of free-form text documents, is typically computationally resource intensive, as well as being time intensive. Without standardised document formats, it can be difficult to achieve consistently accurate results and generate meaningful insights. ESFA needs to build the relevant technical skills in-house to adopt and maintain a new potential Big Data architecture. Data volumes to be stored are growing at a fast pace, and appear in different formats (i.e., structured and unstructured). Therefore, frequent updates in the data infrastructure technologies and tools may be required. Expected Impact www.efsa.europa.eu/publications 57 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence - Better data management and improved transparency and accessibility to data will empower EFSA users to experiment with new advanced analyses that will have a positive long-term impact on EFSA operations and decision-making. - Foundations in place, thanks to a consistent Data Infrastructure, for the effective development of AI algorithms. - Lower maintenance cost and reduction of effort needed. Reduction of IT-Workarounds. Acceleration of the responsiveness to changes. - The Development and adoption of a Data Architecture for Big Data and AI is a foundational element for building the appropriate infrastructure and tools to achieve improved data extraction, integration and standardisation to support the use of NAMs in chemical risk assessment. The following table provides the SWOT analysis for Recommendation 2 SWOT Analysis Strengths - EFSA has already established a partnership with Microsoft Azure, which offers the most cutting-edge cloud platforms in the market. - Microsoft Azure offers various complementary and additional services that can be integrated into the designed Data Infrastructure. Opportunities - Weaknesses - The design of a modern and agile data infrastructure that can lead to the optimisation of the data flows throughout the different data layers from storage to visualisation. Threats Security is essential in a data infrastructure. The development of a data governance framework for EFSA is essential in order to set data policies and controls over the data. Currently, in EFSA, data are stored in multiple silos (i.e., repositories of data that are controlled by one department or business unit and isolated from the rest of an organisation, which makes it hard for users in other part of the organisation to access and use the data). Therefore, to ensure the design of a modern data infrastructure, these data silos should be broken down. - The design of a modern data infrastructure should not be performed independently, and before the adoption of a solid AI governance framework. Otherwise, there would be a risk of not selecting and implementing the correct compliance and security tools. Collaboration with a third-party for the acquisition of new solutions or services that compose a data infrastructure is always associated with some outsourcing risks. For instance, protecting personal and sensitive data (hosted within the data infrastructure) can be a barrier that hinders or delays successful outsourcing. Failure to implement privacy and data protection correctly can result in severe legal, economic, and reputational penalties. Recommendation 3: Implement a Data Governance framework which addresses AI Description and scope As outlined in its 2027 strategy, EFSA aims at becoming an AI-enabled organisation39, and challenges such as data being scattered across the entire organisation in silos and not managed appropriately need to be addressed. Indeed, a foundational element to the adoption of AI solutions is the management and utilisation of a large amount of 39 EFSA, Theme (concept) paper on Artificial Intelligence in risk assessment www.efsa.europa.eu/publications 58 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence structured and unstructured data obtained from multiple sources and in multiple formats. AI encompasses the entire cycle of data capture, data storage, data preparation, and advanced data analytics technologies. Therefore, to successfully adopt AI, EFSA needs first and foremost to ensure that its data is accessible, relevant, reliable, and protected. Within this context, the adoption of a Data Governance framework should ensure that data within EFSA is managed and used correctly and responsibly while minimising data management effort and operational costs, increasing transparency, and minimising risks, leading to better AI-enabled decision making. In this sense, this recommendation does not exclusively focus on the scope of evidence management, but it addresses EFSA as a whole organisation, considering the inputs collected from other EFSA units and the needs identified in other roadmaps, such as NAMs. Indeed, the Development and adoption of a Data Governance framework which address AI is a foundational element for building the appropriate infrastructure and tools to achieve improved data extraction, integration and standardisation, in order to support the use of NAMs in chemical risk assessment. Specific activity A Data Governance (DG) framework essentially describes policies, roles, standards, and metrics needed to continuously improve the use of data that ultimately enables an organisation to achieve its goals. Some of the questions a DG framework should answer are set out in Figure 23 below: Figure 23- Questions that a DG framework answers Figure 24 below shows the structure of an indicative framework for the establishment of DG which can provide the answers to the above questions: Figure 24 - Indicative DG Framework incorporating Tools & Technology, Process & Operating Model and People & Culture. www.efsa.europa.eu/publications 59 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence By answering these questions, DG ensures the quality and security of an organisation's data reservoir since it clearly defines who is responsible for what, and what actions should be taken (using what methods). In addition, addressing these questions would be the basis for a good Responsible AI framework, as data bias and data management are the keystones of the overall ethicality of an AI project (this is covered in more details in horizontal recommendation 5). In the figure below (see Figure 25), a roadmap is set out including potential steps which EFSA should take for the development and, consequently, adoption of a DG framework: Figure 25 - DG roadmap with increasing levels of maturity. Based on the current maturity of the DG development, EFSA should start with Technology-Based DG, focusing initially on the aspects of governance affecting tools, technology and data management before advancing to a more complete DG framework (the regulatory-based governance is an optional step), which would also cover organisational structure, processes and people across the organisation. It is noted that, for organisations like EFSA that are planning to leverage AI, further components need to be taken into consideration and aligned with the DG framework. These are: Development & Operations Management (DevOps) and Trustworthy AI. For DevOps, DG is its backbone, as it sets the processes, structures and roles to ensure that qualitative data are provided for the model development, and that data lineage and continuity is preserved (see Horizontal Recommendation 4). Moreover, setting up a DG framework that reflects the key tenants of Trustworthy AI (see Horizonal Recommendation 5) should contribute to a fairer adoption and use of AI technology. www.efsa.europa.eu/publications 60 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Recommendation Summary Objectives - In general, it is recommended to EFSA to increase the control over its data assets to monitor, control and manage them easily and more effectively, and to provide holistic, and qualitative data when needed. EFSA should consider adopting a detailed data governance plan in order to establish roles, processes and policies for improving EFSA's operational efficiencies. As a parallel consequence for the adoption of a DG framework, EFSA will set the foundations for Trustworthy AI, and will ensure that the AI decision-making process is explainable, transparent and fair. Challenges - The establishment of a DG may require several organisation's units to get involved, hence it can cause misconceptions over data governance. - Third party and other data that the organisation doesn't control. Difficulties with access to data. Fragmented or siloed teams. Expected Impact - An effective DG framework adoption might support EFSA to effectively reduce risks, improve trust, obtain efficiencies and adopt AI in a more trustworthy and transparent way. - An effective DG framework adoption supports a better understanding of the different data types that are available and of their quality, leading to the possibility to reveal and realise the maximum value from data (e.g., identification of new AI projects, or gaps to be bridged). - The Development and adoption of a DG is a foundational element for building the appropriate infrastructure and tools to achieve improved data extraction, integration and standardisation to support the use of NAMs in chemical risk assessment The following table provides the SWOT analysis for Recommendation 3 SWOT Analysis Strengths - Weaknesses Set clear rules for managing the entire data lifecycle (including data capture, data storage, data preparation, and advanced data analytics technologies). - Ensure the accuracy, availability, integrity, and confidentiality of data by defining and following a data governance framework. - Increase data processing efficiency by reducing the lengthy coordination process. - - Opportunities - Threats Increase confidence in data, through qualityassured standard based on the Data Governance Framework. - - www.efsa.europa.eu/publications Consider a time and effort consuming procedure to tailor the Data Governance framework to the specific needs and requirements of EFSA. Resistance to change for such a pervasive action is to be foreseen, change management activities must be considered. 61 Potential difficulties in engaging EFSA stakeholders through new practices and ways of working to support the adoption of a Data Governance Framework. Misalignments among organisation units to establish the Data Governance Framework. DG would evolve with the organisation’s needs, so continuous monitoring and maintenance to adapt to change is required. EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Recommendation 4: Establish Development & Operations Management (DevOps) Description and scope An important component that can help the implementation of a robust Data Governance (DG) framework is the adoption of a Development & Operations Management (DevOps). DevOps is a set of practices, tools, and cultural philosophy that automate and integrate the processes between data science/analysis teams ( development) and IT teams (operations),40 increasing an organisation’s ability to deliver applications and services at high velocity. The function of DevOps is to enable teams to collaborate and communicate more effectively and efficiently, and to ensure continuity and reproducibility in any model and product developed within EFSA. We recommend that EFSA establishes a DevOps practice, and therefore adopts an ecosystem for building, testing, deploying, and maintaining the enterprise data science and machine learning models, as well as other analyses and tools, in production environments. The tangible benefits that DevOps can provide EFSA are included in Figure 26: Figure 26 - DevOps Benefits. Specific activities Our experience has shown that people, processes and technology are all critical elements to establish an effective DevOps practice. Therefore, for a successful DevOps implementation, EFSA should address all these three elements in parallel as follows: • People: Design of an ad-hoc change management program for the whole organisation - or specific department which would promote sufficient training, education programmes communications actions, coaching sessions, and exemplary management, with a strong emphasis on supporting the DevOps delivery team and management; • Processes: Adherence to DevOps design principles; • Technology: Design of an independent modular architecture, with uniform and fit-for-purpose automation software; Automation software for DevOps implementation allow tasks to be performed with reduced human assistance, facilitating the tracking of the model development’s progress. In addition, this software provides quick feedback loops between operations and development teams (data scientists in this case), so that iterative updates can be deployed faster to applications in production. The basic components of these tools include logging (mainly model and experiment parameters), version control and use of dashboards for monitoring activities and performance. Git, Docker, SonarCube and Kubernetes are some of the most widely used open-source DevOps tools. Within the Microsoft Azure environment, Azure DevOps provides services for allowing teams to plan work, collaborate on code development, and build and deploy applications. Azure DevOps supports a culture and set of processes that bring developers, project managers, and contributors together. 40 Under a DevOps model, development and operations teams are no longer “siloed.” Sometimes, these two teams are merged into a single team, where the engineers work across the entire application lifecycle, from development and test to deployment to operations, and develop a range of skills not limited to a single function. www.efsa.europa.eu/publications 62 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Recommendation Summary Objectives The main objectives for Recommendation 4 can be summarised as follows: - It is recommended that EFSA adopts a DevOps, with the aim of building a series of practices and a culture to promote collaboration between development and operations teams within EFSA. - In addition, EFSA should consider establishing a DevOps to increase trust, accelerate model development and deployment, build traceability and continuity, and develop the ability to solve critical issues quickly, and better manage unplanned work. Challenges - Obstacles in establishing an efficient DevOps might be identified within its three main elements, as follows: o o o People: There is a need to build a team with various and advanced skills, such as knowledge of the machine learning model lifecycle, and knowledge of the responsibilities that the data/machine learning/software engineer must have. For this scope, there might be an unwillingness or inability to broaden scope and skill-sets to those required. Processes: Different employees may end up working for multiple teams, and teams may become too large when trying to merge development and operations processes. Technology: There might be issues related to interdependencies in responsibilities and Technology interoperability and integration. Expected Impact - Elimination of manual hand-offs to reduce time for deployment and human errors Modernisation of systems to avoid bottlenecks Increase in transparency and improved collaboration Cost reduction, since no work is duplicated and errors are reduced The following table provides the SWOT analysis for Recommendation 4 SWOT Analysis Strengths - Weaknesses EFSA already uses Microsoft Azure, which provides the Azure DevOps. Open-source tools are available, which can kickstart the DevOps activities with minimal investment. - Opportunities - The organisational challenges with bringing the teams together may cause delays or resistance to adoption. Threats Improve collaboration, reduce cost, reduce errors and develop an all-around better culture. Focusing only on the adoption of a DevOps tool instead of tackling the issue from a people perspective might lead to failure. Recommendation 5: Adopt a Trustworthy AI Framework Description and scope With a view to EFSA's vision of scaling responsible and human-centric AI by 2027, the adoption of a Trustworthy AI framework, which can ensure that AI challenges are addressed from an ethical, social and legal point of view, is highly recommended. As we saw in section 3.1, the adoption and use of AI poses several challenges and risks to EFSA regarding ethics and compliance with the data protection Regulation (EU GDPR), which need to be addressed by EFSA to establish trust with their stakeholders. In fact, these risks and challenges represent great societal concerns for EFSA's main stakeholders (e.g. NGOs). Furthermore, it is worth mentioning the guidelines that the European Commission set in the Proposal for a Regulation of The European Parliament And Of The Council Laying Down www.efsa.europa.eu/publications 63 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Harmonised Rules On Artificial Intelligence (AI Act) 41, which aim to strengthen AI update across private and public organisations, while addressing AI risks, in order to ensure that AI systems are based on rules that safeguard the public sector's operations and fundamental rights. To take the draft of the AI Act into consideration when developing the Trustworthy AI framework would ensure that the AI systems adopted by EFSA are safe and legally compliant, and respect fundamental rights, as well as European Union values. In addition, by adopting the framework, EFSA would increase transparency towards both its internal and external stakeholders about how AI and data are used, what are the fundamental decisional processes that are followed, and how the potential risks are identified and addressed. Trustworthy AI framework design and structure The suggested Trustworthy AI framework consists of a series of rules, governance mechanisms and tools aimed at guiding the design, development, and deployment of trustworthy AI tools. When efficiently adopted, the Trustworthy AI framework empowers organisations to work with agility and enhance trust towards the use of AI solutions. As organisations deploy AI to automate existing or new workflows (either with the aim of reducing time and/or effort consumption and/or increase efficiency and results), the framework enables users to trust AI outcomes at every step of the AI lifecycle (i.e., design, development, deployment, operating and monitoring phases), and also ensures enhanced transparency on AI use to externals stakeholders. In order to do so, the Trustworthy AI framework considers the requirements proposed in the draft of the AI act and translates42 them into recommendations and suggested actions (see Table 22 below). The AI Act presents a balanced and proportionate regulatory approach to AI, that tries to provide minimum necessary requirements to address the risks and problems linked to AI through a risk-based approach. Indeed, the AI Act describes four risk levels for every AI solution and provides different types of requirements for each of them. Table 12 below illustrates how the most important requirements per risk level of the proposed AI Act map towards the Recommendations for EFSA presented in this document, as well as the corresponding specific actions. Table 12 - Mapping between AI Act requirements and recommendations and actions for EFSA. AI Act Risk Level AI Act Requirements # Recommendation Suggested actions Minimal or no risk Create and implement a Code of Conduct, which may also include voluntary commitments related, for example, to environmental sustainability, accessibility for persons with disability, stakeholders’ participation in the design and development of AI systems, and diversity of development teams. Recommendation N.1 Adopt a trustworthy AI framework Action 1 “Develop a Code of conduct” Limited Risk Ensure transparency for system which interact with humans, detect emotions or determine association with (social) categories based on biometric data, or (iii) generate or manipulate content (‘deep fakes’). No specific recommendation as these requirements are very specific and asked only for some type of AI applications. Establish and processes. Recommendation N.1 Adopt a trustworthy AI framework High Risk implement risk management Adopt data and data governance requirements on training, validation and testing data Recommendation N. 3 Governance framework43 Action 2 “Create a Risk Management Framework” Data 41 European Commission, Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence and Amending Certain Union Legislative Acts, (2021) available at https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52021PC0206 42 It should be noted that the fourth risk concerns applications with an unacceptable risk, which are not allowed to be used 43 A data governance framework can be considered as a methodology to implement specific principles of Trustworthy AI concerning data management. For more information, see Recommendation 3 “Data Governance framework”. www.efsa.europa.eu/publications 64 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence AI Act Risk Level AI Act Requirements # Recommendation Establish documentation and design features (traceability & auditability) • • • logging Ensure appropriate certain degree of transparency and provide users with information Ensure human oversight Ensure robustness, accuracy and cybersecurity Suggested actions Recommendation N. 4 Establish Development Operations Management (DevOps)44 Recommendation N.1 Adopt a trustworthy AI framework Action 2 “Create a Risk Management Framework” Recommendation N.4 Establish Development Operations Management (DevOps) 45 As the table shows, some requirements from the AI Act are addressed by Recommendation N. 3 Data Governance framework and Recommendation N. 4 Establish Development Operations Management (DevOps), because they already implement some specific governance measures for those requirements. For example, the Data Governance framework as described in Recommendation N. 3 already helps EFSA in the implementation of some AI principles such as Privacy and Data Governance or Transparency, especially in terms of Traceability and Accountability. 46 In addition, when developing this, additional questions related to ethical issues could be considered, such as: o o o o o What phenomena are represented in the data? How was the data collected? Is the data balanced? Does the data include sensitive information or proxies? Does the data hide some cultural or historical bias? Towards which groups? Similarly, the adoption of Development & Operations Management (DevOps) (Recommendation N. 4) will address several AI principles (such as traceability and auditability) and will integrate some requirements of the AI Act specific for the development, deployment, operating and monitoring phases of AI solutions. In addition to the actions stemming from the AI Act, a third action has been added to the Trustworthy AI framework recommendation, which aims at ensuring an efficient adoption of the framework itself. This last action concerns internal skills, capabilities and roles. Specific activities The complete list of the suggested actions under Recommendation 5 - Adopt a Trustworthy AI Framework, which EFSA should follow to adopt a robust and efficient Trustworthy AI framework consists of: • • • Action 1 “Develop a Code of conduct” Action 2 “Create a Risk Management Framework” Action 3 “Organise internal skills, capabilities and roles” Action 1 “Define organisation principles by developing a Code of Conduct”: Firstly, it is recommended that the organisation adopts a Code of Conduct (CoC), which is a statement of principles aimed at guiding its own conduct. These principles can serve both as affirmations of values and as the basis of an internal accountability mechanism for the organisation. Codes of conduct usually focus on ethical requirements and recommendations that go beyond what the law requires. Indeed, they can contribute significantly to an organisation’s ability to deal with difficult ethical issues, especially if the code is developed in a process with broad participation in the organisation. For this reason, 44 Development Operations Management (DevOps) can be considered as another methodology to implement specific principles of Trustworthy AI concerning model management. For more information see Recommendation 4 Establish Development Operations Management (DevOps) 45 Development Operations Management (DevOps) can be considered as another methodology to implement specific principles of Trustworthy AI concerning model management. For more information see Recommendation 4 Development Operations Management (DevOps) 46 See the Requirements of Trustworthy AI from the Ethics Guidelines for Trustworthy Artificial Intelligence prepared by the High-Level Expert Group on Artificial Intelligence https://ec.europa.eu/futurium/en/ai-alliance-consultation/guidelines/1.html www.efsa.europa.eu/publications 65 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence EFSA’s board members and executives should participate in a collaborative process for establishing the CoC in order to ensure that the entire organisation is pushing towards – and committed to – the same direction. After the code has been adopted, EFSA might consider establishing an ethical committee to further contribute to the organisation’s ethical culture and its ethical competence (see Action 3). As a starting point for the development of a more detailed Code of Conduct, EFSA might consider, as a relevant example, the Ethics Guidelines for the Trustworthy Artificial Intelligence developed by the High-Level Expert Group on AI (HLEG)47. This document defines Trustworthy AI: it consists of three aspects, which should be met throughout the AI system's entire life cycle: • • • lawful - respecting all applicable laws and regulations ethical - respecting ethical principles and values robust - from a technical perspective, while taking into account its social environment. Based on these three aspects, four main principles have been established to ensure that AI systems are developed, deployed and used in a trustworthy manner: Respect for human autonomy, Prevention of harm, Fairness, Explicability. To realise and implement these principles, the HLEG defined 7 key requirements, which are now considered as a standard de facto48 across the world for how to implement Trustworthy AI. These requirements are: Table 13 - Map of the HLEG’s Requirement with short description HLEG’s Requirement Short description Human agency and oversight AI systems should empower human beings, allowing them to make informed decisions and fostering their fundamental rights. At the same time, proper oversight mechanisms need to be ensured, which can be achieved through human-in-the-loop, human-on-the-loop, and human-in-command approaches. Technical robustness and safety AI systems need to be resilient and secure. They need to be safe, ensuring a fall-back plan in case something goes wrong, as well as being accurate, reliable and reproducible. That is the only way to ensure that also unintentional harm can be minimised and prevented. Privacy and data governance Besides ensuring full respect for privacy and data protection, adequate data governance mechanisms must also be ensured, taking into account the quality and integrity of the data, and ensuring legitimised access to data. Transparency The data, system and AI business models should be transparent. Traceability mechanisms can help achieving this. Moreover, AI systems and their decisions should be explained in a manner adapted to the stakeholder concerned. Humans need to be aware that they are interacting with an AI system, and must be informed of the system’s capabilities and limitations. Diversity, nondiscrimination and fairness Unfair bias must be avoided, as it might cause multiple negative implications, from the marginalisation of vulnerable groups, to the exacerbation of prejudice and discrimination. Fostering diversity, AI systems should be accessible to all, regardless of any disability, and involve relevant stakeholders throughout their entire life circle. Societal and environmental well-being AI systems should benefit all human beings, including future generations. It must hence be ensured that they are sustainable and environmentally friendly. Moreover, they should take into account the environment, including other living beings, and their social and societal impact should be carefully considered. Accountability Mechanisms should be put in place to ensure responsibility and accountability for AI systems and their outcomes. Auditability, which enables the assessment of algorithms, data and design processes plays a key role therein, especially in critical applications. Moreover, adequate and accessible redress should be ensured. 47 High-Level Expert Group on Artificial Intelligence (2019) “Ethics Guidelines for trustworthy AI.” European Commission. Available at: https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai 48 Fjeld J., Achten N., Hilligoss H., Nagy A., Srikumar M., Principled Artificial Intelligence: Mapping Consensus in Ethical and Rights-Based Approaches to Principles for AI, Berkman Klein Center Research Publication, 2020 www.efsa.europa.eu/publications 66 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Another relevant example for EFSA to take inspiration from is given by the general Code of Conduct 49 developed by ETAPAS50, a project funded by the European Commission, which includes ten fundamental principles to guide the correct implementation of Disruptive Technologies (i.e., Artificial Intelligence, Big/Open Data, Blockchain, Augmented and Virtual Reality, Internet of Things and Robotics) by Public Sector organisations. These principles are set out in Table 14 below. Table 14 - List of ETAPAS principles with short description Principle Short description 1 Environmental sustainability It is important to use Disruptive Technologies (DTs) conforming to the principles of environmental sustainability. The principle considers the production processes for the hardware and the materials used to produce the new technology, including possible new materials and their impact on the environment. Biodiversity is a precondition for ecological resilience, and it must always have an important role in the evaluation of environmental effects of new technologies. 2 Justice, equality, and the rule of law Equal treatment and the rule of law are fundamental values from which public administration should never deviate. The introduction of DTs should promote equality of access and opportunity to everyone, and aim at eliminating possible biases towards specific groups (e.g. women, minorities, etc.) due to low quality or missing training data for algorithms used (e.g. AI, machine-learning or decision algorithms). The input data for algorithms needs to be carefully selected and evaluated in order to make sure that discrimination or other undesired effects are not introduced. Automated decision-making that impacts on individuals should only be used when there is reason to be confident that the algorithm does not discriminate against any group of individuals. 3 Transparency and explainability Individuals who are affected by a public decision based on automated data processing should have access to clear information that a layperson can understand, both on how the decision was made (transparency), and on its justification (explainability). Information must also be presented in an accessible way for disabled and elderly people. 4 Responsibility and accountability The public sector is subject to strict principles of accountability, and blaming a machine, an algorithm or a decision support system for a decision affecting individuals or society is not a feasible solution. There must always be sufficient human oversight and control of automatic decision-making to ensure that human decision-makers can be held accountable. Adequate procedures must be in place for communication with individuals, and for proper investigation and mitigation of problems arising from the use of DTs. 5 Safety and security Before the introduction of DTs, the risks of both intentional and unintentional harm should be carefully evaluated. The technological and organisational measures need to prevent the harmful effects DTs could have on individuals and any other specific security risks that may appear (e.g. hacking, data misuse, data manipulation). 6 Privacy Best-practice methods for data protection, including deletion mechanisms, should be used, and these methods should be regularly updated. With access to large quantities of data, unintended harmful consequences for individuals might occur. Thus, use of such data should be strictly regulated, and efficient measures to prevent and discover unauthorised use should be implemented. 7 Building an ethical culture involving employees The participation in the creation of an ethical culture at the workplace is essential for the functionality and credibility of the public sector. The DTs should not be invasive forms of control of workplace behaviour, but tools to relieve public servants of routine work and make better use of their competences. The DT introduction must be decided in a participative and 49 The Code of Conduct (CoC) of ETAPAS is available on the project website at the following link: https://www.etapasproject.eu/resources/deliverables/ 50 ETAPAS (Ethical Technology Adoption in Public Administration services) has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 101004594 www.efsa.europa.eu/publications 67 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Principle Short description co-creative process involving both the public servants and their organisations. Upskilling and reskilling, including for the ethical use and management of new technologies, should be part of the working conditions. 8 Retaining human contacts Ensuring the wellbeing of all residents must be a priority of the public sector, also in the adoption of new technologies. While the new technologies might negatively impact the social contacts, they should be used in ways that do not weaken the social networks and make residents feel more isolated. The introduction of alternative human-based modes of contact should be considered whenever needed to avoid the possible negative effects of DTs adoption. 9 Ethical publicprivate cooperation The public sector should have enough competences of its own and ability to control and review private sector involvement to ensure that such cooperation works efficiently in the public interest. Private sector performance and accountability should be ensured through appropriate contractual protections. Artificial decision support systems in the private sector that are subject to public supervision must be sufficiently transparent to make sure that efficient supervision can be performed. 10 Continuous evaluation and improvement The social effects of disruptive technologies are difficult, often impossible, to foresee. When the technology is introduced on a larger scale, evaluation should continue in order to detect (positive and negative) effects that may not have been discovered in the small-scale trials. These evaluations should include the perspectives of residents and employees, particularly groups at risk of exclusion or discrimination from technologies. Adjustments, improvement and when necessary, replacement of technologies should be made whenever needed. Action 2 “Create a Risk Management Framework”: Once the Code of Conduct is drafted, it will be necessary to develop practical guidelines which can guide the trustworthy adoption and monitoring of AI solution. Indeed, principles within the Code of Conduct should be translated into a Risk Management Framework which would provide a detailed mapping of all the risks to which EFSA might be exposed when adopting an AI solution and governance mechanisms to control and mitigate such risks. Starting from the good practices of risk management in various sectors (e.g. financial services where these activities are very mature), it is important to define a taxonomy of possible risks to which EFSA could be exposed when adopting and using AI. For example, within the above-mentioned ETAPAS project, a risk framework 51 has been developed providing a detailed mapping of all the risks to which public bodies might encounter when adopting DTs. In particular, it identifies eight risk categories including 35 risks, which can be traced back to the 10 ethical principles of the Code of Conduct, as follows: a. Risks concerning direct interaction with humans, such as the risks of physical and psychological harm; b. Legal Risk, including liability risk and law infringement risk; c. Governance risks, which comprises all the risks arising from mismanagement, weak supervision and lack of standard procedures; d. Enhanced inequality and discrimination, which covers risks related to possible perpetuation of discrimination with DT-based solution and the unequal access to and benefit from the DTs, due to economic, cultural and social differences; e. Errors and misuse, including risks emerging from the DT lack of robustness, biased performances or security branches, as well as the risk of the DT's being used as an autonomous weapon, for malicious surveillance purposes or for spreading disinformation while limiting the decision-making power of citizens; 51 For more information, see the contribution to the OSAI21 workshop at the following link: https://www.unive.it/pag/fileadmin/user_upload/progetti_ricerca/osai/img/grafica/OSAI21_paper_5.pdf www.efsa.europa.eu/publications 68 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence f. Security and data protection risks, such as the risk of the DT being hacked to get information access or to change its behaviour; g. Unsustainable use, stemming from high level of energy consumption or wide use of environmentally unsustainable materials; h. Workplace issues, including job displacement and lack of competences by public sector employees. Once all of the risks have been identified, different governance mechanisms could be put in place, such as the identification and definition of controls for each risk in a post-hoc evaluation of AI solutions, or the adoption of an ex-ante assessment to drive the development of AI solution from the design phase. From our analysis, the latter would suit EFSA’s context better. The development and adoption of an assessment to evaluate and measure the AI impacts, and determine potential risks and mitigations actions would constitute a practical method to integrate the risk management framework within the entire AI lifecycle. The assessment should aim at ensuring an appropriate and certain degree of transparency, and provide users with information, human oversight, robustness, accuracy and cybersecurity, while adopting the AI solutions. Usually, these assessments have a twofold objective: provide practical actions and guidelines to mitigate the possible risks, while assessing and monitoring the residual risk. Also in this case, EFSA might take into consideration existing best practices, such as the Assessment List for Trustworthy AI (ALTAI) developed by the High-Level Expert Group on AI 52. The ALTAI translates the principles and the HLEG trustworthy AI requirements into an accessible and dynamic checklist aimed at implementing such principles in practice. It consists of a series of questions - both technical and non-technical – which assess whether the development process and the AI system address the selected risk. In the Figure below, an example of questions referring to the principle of human agency and oversight is detailed. Figure 27 - ALTAI example of questions which refer to the human agency and oversight principles In addition, referring once again to the ETAPAS project, an assessment – called Indicator framework - has been built as a way to track and measure relevant risks and their mitigation actions. This tool was developed specifically for public sector organisations that want to adopt Disruptive Technologies in a responsible manner. The ETAPAS Indicator Framework, made up of a set of indicators, has been defined and structured as a checklist. Within the framework, two types of indicators have been selected: those assessing whether the risk is relevant for the context of application (risk indicators), and those that evaluate how much the risks were mitigated based on specific mitigation actions (mitigation indicators). If the risk is present, then the mitigation questions should be addressed too. Considering Figure 28 below, an example of a risk question could be: “Does the DTA (Disruptive Technology Application) interact with the decision-making process of human end-users?”. Then, mitigation follow52 https://digital-strategy.ec.europa.eu/en/library/assessment-list-trustworthy-artificial-intelligence-altai-self-assessment www.efsa.europa.eu/publications 69 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence up questions are listed which suggest actions of a different type: e.g., “Will the relevant personnel be able to assume control where necessary?”; or “How do users feel when taking decisions based on the DT application?”. Figure 28 - ETAPAS Indicator framework design and structure. Action 3 “Organise internal skills, capabilities and roles”: Designing and deploying trustworthy AI systems should be an organisation-wide effort. It requires sound planning, cross-functional and coordinated execution, employee training for the new required skills and capabilities, and significant investment in resources to drive the adoption of responsible AI practices. Indeed, employees need to be trained to understand how risk manifests in their contextual interactions with AI systems and, more importantly, how to identify, report and mitigate them. Significant effort should be dedicated to tailor the training contents to the different roles in the organisations, in order to provide the correct level of information and upskilling of all of the people53. In particular, the Code of Conduct should be well known across the whole organisation to actually drive the behaviour of each person. In addition, investment in resources for the establishment of new roles and/or integration of new responsibilities in the current organisational structure would be fundamental. For example, the establishment of an ethical committee might be considered to drive the development and monitoring of EFSA’s Trustworthy AI Framework. This ethical board might become the central organism for the ethical decision-making concerning EFSA’s AI solutions, responsible for the resolution of possible ethical issues and tensions. These considerations should be taken into account within Recommendation N. 6 Develop data science competencies & design behaviour change mechanisms. Within this context, different groups of stakeholders play different roles in ensuring that the practical guidelines are met54. Some examples are provided the table below (Table 15) 55: Table 15 - Example of organisational governance. Role Management and Board Compliance/Legal Short description • Top management discusses and evaluates the AI systems’ development, deployment or procurement, and serves as an escalation board for evaluating all AI innovations and uses, when critical concerns are detected. It involves those impacted by the possible introduction of AI systems (e.g. workers) and their representatives throughout the process via information, consultation and participation procedures. • The corporate responsibility department monitors the use of the assessment list and its necessary evolution to meet the technological or regulatory changes. It updates the 53 For example, Recommendation 6. Development of data science competencies and design of behaviour change mechanisms should also consider ethical AI related skills. 54 For information on the technical profiles needed to implement the Trustworthy AI framework, see Recommendation 6 “Data Science Competencies & behaviour change mechanism” 55 This example refers to the proposed organisational governance for the ALTAI, available herehttps://ec.europa.eu/futurium/en/ai-alliance-consultation/guidelines/2.html www.efsa.europa.eu/publications 70 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence department/Corporate responsibility department Product and Development equivalent Service or standards or internal policies on AI systems and ensures that the use of such systems complies with the current legal and regulatory framework and to the organisation's values. • The Product and Service Development department uses the assessment list to evaluate AI-based products and services, and logs all the results. These results are discussed at management level, which ultimately approves the new or revised AI-based applications. Quality Assurance • The Quality Assurance department (or equivalent) ensures and checks the results of the assessment list, and takes action to escalate an issue to a higher level if the result is not satisfactory, or if unforeseen results are detected. HR • The HR department ensures the right mix of competences and diversity of profiles for developers of AI systems. It ensures that the appropriate level of training is delivered on Trustworthy AI inside the organisation Procurement • The procurement department ensures that the process to procure AI-based products or services includes a check of Trustworthy AI. Day-to-day operations • Developers and project managers include the assessment list in their daily work, and document the results and outcomes of the assessment. It is important to note that the success of introducing the Trustworthy AI Framework in EFSA (and any other organisation) is facilitated by smooth and incremental changes to the current organisation. For this reason, we recommend defining the target processes, roles and responsibilities based on a thorough assessment and analysis of the as-is situation. Recommendation Summary Objectives The main objectives for Recommendation 5 can be summarised as follows: - In order to adopt the AI system in a responsible way, it is suggested that EFSA adopts a Trustworthy AI framework that addresses the challenges around AI from an ethical, social and legal point of view. Challenges The main challenges for Recommendation 5 can be summarised as follows: - The establishment of a Trustworthy AI framework may require several of the organisation's units to get involved, thereby possibly causing misconceptions over the framework. Converting principles of the Code of Conduct into practice can be a difficult task. The technical complexity and scale of people and process change might be underestimated. The organisation could struggle to scale the Trustworthy framework within the lifecycle of the AI system. Expected Impact The main expected impact for Recommendation 5 can be summarised as follows: - An effective Trustworthy framework adoption might support EFSA to effectively improve trust, reduce risks, obtain efficiencies, and adopt AI in a more trustworthy and transparent way - Once AI techniques are applied in NAMs for decision support in risk assessment, the adoption of a Responsible AI framework will also be of paramount importance. The following table provides the SWOT analysis for Recommendation 5. SWOT Analysis Strengths www.efsa.europa.eu/publications Weaknesses 71 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence - Set clear rules and policies for the adoption of the Trustworthy AI Ensure the trustworthiness, transparency, and fairness of the outcomes of AI solutions - Opportunities - - Time and effort-consuming procedures due to the organisation alignment to proper AI principles Threats Increase confidence in the outcome of the AI solutions through transparency and a fairnessassured standard based on the AI governance Framework Anticipate some requirements of the future European AI regulation - Misalignment among organisation executives while adopting AI principle Misalignment of different technical units while adopting AI principle within the AI lifecycle Difficult comprehension of terminology used in the Trustworthy AI framework among BU of the organisation due to lack of new skills Recommendation 6: Expand data science competencies & design behaviour change mechanisms Description and scope In order to efficiently adopt AI for evidence management and to support the actions around the previous 5 recommendations, it is necessary to expand data science competences (via recruitment and upskilling) such as natural language processing (NLP), text mining, ethical AI and DevOps. The introduction of innovative tools and new collaborative ways of working might require training and upskilling of EFSA internal staff. In addition, EFSA might consider developing behaviour change mechanisms, which would facilitate the transition of EFSA towards an AIenabled organisation. In more detail, EFSA competencies needed to effectively adopt and use the identified AI tools are: • NLP and text mining: Natural language processing (NLP) comes from a combination of computer science, information sciences, AI, and linguistics. Therefore, current data science skills within EFSA are not sufficient for moving towards the use of NLP. The necessary skills of a NLP data scientist/engineer are: o o o o o o o • Knowledge of machine learning and deep learning frameworks and libraries Knowledge of Big Data frameworks like Spark and Hadoop The ability to understand text representation techniques, algorithms, statistics Text classification & clustering skills Programming skills – Python, Java and/or R Familiarity with Syntactic & Semantic Parsing Experience with ontologies and knowledge graphs Domain Knowledge Expertise: While not a specific prerequisite, a general understanding of the field of study in which AI solutions are applied would be beneficial to generate higher value for EFSA. A data scientist with domain-specific knowledge can better support the transition towards the adoption of AI technology, acting as a point of contact between the users and the AI-based tools implemented. Because of the difficulty of recruiting data scientists with expertise in the subjects EFSA is working on, it might be beneficial to consider recruiting individuals with NLP engineering skills to undertake research projects (e.g., PhDs in collaboration with academic institutions) within specific EFSA departments in order to acquire the domain expertise and apply their technical skills in a more targeted way. • Skills on Trustworthy AI: To move towards the adoption of the Trustworthy AI, it is fundamental that EFSA data scientists acquire technical skills about this. These skills include: o Testing for bias in the data, model, and human use of AI algorithms to improve fair treatment across groups. www.efsa.europa.eu/publications 72 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence o o o Adding transparency, explainability and provability to the modelling process to improve human understanding of the model outputs. Improving security and robustness of AI through rigorous validation, continuous monitoring and maintenance, verification and adversarial modelling (this is addressed via the DevOps recommendation). Understanding of moral and systemic implications of using AI during development and adoption (in collaboration with business users). Specific activities In order to take full advantage of its data and to better understand the logic of the results when using AI-based tools in the evidence management process, EFSA data science capabilities should be upgraded. The following five steps (Figure 29) might help EFSA to achieve this aim, and, at the same time, create the basis for implementing the mechanisms regarding the behaviour change. Figure 29 - Steps for Data Science Competences Development. Designing behavioural change mechanisms It is important that - along with the introduction of new tools, practices, structures and ways of working -, EFSA undertakes targeted change management activities in order to establish the behavioural changes that support the success of the change. In the light of this, for the implementation of any project related to both the horizontal and vertical recommendations, we recommend EFSA to include dedicated change management actions to ensure the successful adoption and the establishment of the appropriate behavioural changes Figure 30 outlines the principles that should be taken into consideration in order to engage EFSA staff in the new ways of working. Figure 30 - Six principles of People-centred Change and Adoption. www.efsa.europa.eu/publications 73 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence On the basis of the above-mentioned principles, the four pillars of a change management approach are shown below. Figure 31 - Six pillars of change management approach. Recommendation Summary Objectives - - EFSA should consider expanding data science competencies, either through new hires or upskilling of existing personnel, mainly by introducing NLP and text mining skills. These new competencies will support the efficient adoption and use of AI for evidence management in risk assessment, as well as the generation of new ideas on how to manage and maximise value from data. EFSA should consider change management actions which would enable critical behaviours supporting cultural transformation and adoption of AI technology. Challenges - Data science competencies are comprised of different specific skills including data preparation, modelling and analysis, and visualisation and communication, as well as knowledge of NLP and text processing skills. Within this context, the challenge consists in finding the perfect formula that would ensure the correct adoption and maintenance of AI solutions. - Before recruiting, there is a need to accurately examine the current competences of internal technical profiles (e.g., solutions engineer, data architect, data scientist etc.), identify the gaps and, if deemed necessary, design a personalised training programme or curriculum to upskill them. www.efsa.europa.eu/publications 74 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Recommendation Summary Expected Impact - The development of data science competencies and behavioural change mechanisms might support EFSA in gaining autonomy with regard to the development, customisation (especially for those solutions with a low technology readiness requiring specific domain expertise) and, mostly, monitoring of AI tools. In addition, considering the low level of readiness of several solutions to automatise specific activities within the evidence management process, developing these competences and mechanisms would support EFSA in the potential development or co-development of AI solutions, and their customisation to EFSA needs. The following table provides the SWOT analysis for Recommendation 6. SWOT Analysis Strengths - Weaknesses EFSA has a sufficient number of experienced staff in the field of Big Data and Data Analytics. EFSA is willing to invest both in the data infrastructure and data governance framework, which play a key role in guiding the development of data science competencies and cultural change of the organisation. - - Opportunities - 4.2. There is not an established and clear way for providing rapid access to the data, while at the same time ensuring a consistent data governance. As a result, this slows down the rapid evolvement of data science practitioners. The organisation structure is not currently designed in a way that it can leverage the collaboration and full potential of the key data science capabilities. Threats There are still many use cases of the evidence management process that are not associated with any commercialised AI-based solution. The combination of advanced data science capabilities and domain specific knowledge within EFSA can be the foundation for the potential internal development or the co-development of AI solutions. - One risk might consist in the misinterpretation of data due to lack of training/domain understanding of data analysts, losing opportunities for high value creation. - Regardless of the data science competencies developed, there is the risk to draw poor conclusions and wrong associations due to possible misunderstanding and use of data. Not enough work in niche areas (e.g., NLP) may be available for specialists to be fully utilised. Vertical Recommendations In the previous section, a series of horizontal recommendations have been developed to enable a smooth and effective adoption of customised AI-based solutions, and to allow EFSA to reach its Strategy 2027 objectives. 56 In fact, horizontal recommendations unlock the possibility to adopt AI in the use cases along the evidence management process (more details on the relations between horizontal recommendations and adoption of AI in use cases are outlined in section 5.2.3). The vertical recommendations that we present in this section make the last step, by associating the prioritised use cases within the evidence management process with one of the 3 possible vendor selection scenarios, following the methodology presented in section 3.3.3. More specifically, for each of the 10 prioritised use cases (resulting from the prioritisation exercise reported in section 3.3.1), the optimal approach to either procuring or developing AI solutions to be adopted by EFSA (vendor selection scenario) with its specific procedures and conditions, is suggested based on its assessment in terms of 56 EFSA, Theme (concept) paper on Artificial Intelligence in risk assessment. www.efsa.europa.eu/publications 75 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence number of commercialised solutions - identified in the market scouting - and coverage level - through the coverage assessment. Table 16 below reports the results of this exercise, linking each prioritised use case of evidence management to the corresponding vendor selection scenario. According to the methodology, prioritised use cases with a high number (at least 3) of available/commercialised solutions (high technology readiness level) and associated to a high coverage level (higher than 90%) are placed into scenario 1, which considers the procurement process as the optimal procedure to purchase off-the-shelf solutions. Alternatively, for the use cases with a high number (at least 3) of available/commercialised solutions (high technology readiness level) and associated with medium level of coverage (between 90% and 60%), scenario 2 is recommended. Within this scenario, procurement or procurement of innovative solutions (PPI), involving a Proof of Concept to test the solutions feasibility and performance, should be considered as the optimal collaboration model for adopting the AI solutions. Finally, use cases with either few available identified solutions (less than 3), no solutions available or with a low level of use case coverage (< 60%) are associated with scenario 3. This scenario suggests launching a grant, contest or PCP as the most suitable collaboration models to finance R&D activities, aimed at developing and implementing AI-based innovative solutions that respond to needs and resolve problems that are not addressed by solutions available in the market. Table 16 below provides an overview of the three AI adoption scenarios in relation to the 10 prioritised use cases in the evidence management process, also listing the potential partners and related AI solutions. www.efsa.europa.eu/publications 76 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Table 16 – AI Adoption Scenarios and connected Prioritised Use Cases of the Evidence Management. Prioritised Use Case Data Collection Terminology Assessment No of available Solutions (>= 3 or < 3) High Market Solutions Coverage Vendor Management Scenario High Scenario 1: Procurement Procedure to buy off-the-shelf and advanced AI solutions SLR Keywords Identification High High SLR Relevant Abstract Screening High High Organisation Solution Octopai Uninova Seven Past Nine Universite de Tours Universiy of Science Technology Bannu Foundation for Research Technologies - Hellas Arab Academy for Science University of Manchester Scipilot SGS Digicomply Iris.ai Sciome Risklick AG EMA Sciome Evidence Partners Seven Past Nine Octopai Research project Seven Past Nine Heterotoki Homogeneous Mapping University of Manchester SLR Literature duplication De- www.efsa.europa.eu/publications High High 77 Rayyan AMNH VKM Eppi-centre Pico Portal Sciome and and Ontology VisTA Word Encoding Method HYPHEN Syras Search Engine Results Page Iris.ai Swift Review Risklick AG Internal solution Swift Review DistillerSR Seven Past Nine Systematic review workload reduction through certaintybased screening Rayyan Colandr Internal solution Eppi-Reviewer Pico Portal Swift Review EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Prioritised Use Case No of available Solutions (>= 3 or < 3) Market Solutions Coverage Text Summarisation High High SLR Literature Findings Clustering High High Quality Improvement of the Scientific Output Expert Identification Pool www.efsa.europa.eu/publications High Low Medium Low 78 Vendor Management Scenario Scenario 2: Procurement or Innovation procurement (PPI) to develop a PoC (Proof of Concept) and integrate AI solutions or to integrate the solution Scenario 3: grant for research and development services or PCP to develop, test and integrate the solutions or Organisation Solution Evidence Partners Risklick AG Iris.ai ASReview Research Screener Robot Reviewer EUIPO CDT Iris.ai Innovamol Consulting Eppi-centre Scipilot SGS Digicomply Sciome Evidence Partners Risklick AG Creme Global LibrAlry Ginger Grade Pro DistillerSR Risklick AG Iris.ai ASReview Research Screener Robot Reviewer Internal solution Internal solution Iris.ai Innovamol Eppi-Reviewer Syras Search Engine Results Page Swift Review DistillerSR Risklick AG Creme Global LibrAlry Ginger Software Grade Pro ProWritingAID ProWritingAID Pomato Department of Scientific Research, Visibilia SP Ramrao Adik Institue of Technology Pomato Job Reccomendation Job Recommender System EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Prioritised Use Case No of available Solutions (>= 3 or < 3) Market Solutions Coverage Vendor Management Scenario Contest to develop, test and validate innovative solutions Expert Selection High Low Organisation Solution University of Electronic Science and Technology of China Textkernel MosaickTrack Springer Nature Ideal Pomato Leoforce Department of Scientific Research, Visibilia SP Ramrao Adik Institue of Technology University of Electronic Science and Technology of China Machine Learned ResumeJob Matching Texkernel MosaickTrack Reviewer Finder AI for Recruiting Pomato Arya Agriculture and Agrifood Canada Comparable Appraisals Detection www.efsa.europa.eu/publications Low Medium Wageningen University & ResearchInformation Technology University of San Francisco 79 Job Recommendation Job Recommender System Machine Learned ResumeJob Matching Exploring Identifiers of Research Articles Related to Food and Disease Using Artificial Intelligence Automatic Systematic Literature Review Critical Appraisal Tools EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence The following tables provide a SWOT Analysis to identify strengths, weaknesses, opportunities, and threats associated with the adoption of AI technology for each prioritised use case that falls under the respective scenario: Vendor Management Scenario 1 – Procurement Swot Analysis – Data Collection Terminology Assessment Strengths - - - Weaknesses There are many available solutions related to this use case that EFSA can evaluate through an effective and efficient procurement process to ultimately choose the best solution in terms of quality and costs. The cost of buying a commercialised tool that need to be customised based on EFSA needs and requirements is less than the amount spent on investing in research and development for a tool/service that does not exist yet, or that is in its initial level of technology readiness. The integration of an AI solution for this use case would have a low impact in the organisation, and is not likely to face resistance to adoption by EFSA staff. Opportunities - - - Considering that the majority of the commercialised AI tools are developed for a general purpose, EFSA may need a unique solution tailored to its needs and context. However, most of the vendors have ensured their solution is highly customisable, even though it may require additional costs and time for implementation. Threats The adoption of AI technology would automate the alignment of the content of data collected to EFSA vocabulary, leading to time/efficiency gains and quality improvement. The PERA team suggested that an AI tool that will be able to evaluate how well an interview has been transcribed into text format could really be a benefit in terms of time and workload reduction. PERA has its own database of texts that have already been transcribed, and therefore this can be used as “training data” for the AI tool, and to further increase its accuracy. Modifying slightly the use case of Data Collection Terminology Assessment would help address the challenge, since the logic of assessing the text format and terms is the same. - There is a potential risk of incurring in unexpected costs if the compatibility of the AI solutions within EFSA’s ecosystem is not properly assessed and ensured. There is a potential risk of a low degree of solution/vendor flexibility in case of interim changes of need (scope, requirements), but this can be mitigated by adopting an agile project management approach. Swot Analysis – SLR Keywords Identification www.efsa.europa.eu/publications 80 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Strengths - - Weaknesses Most of the tools have a good accuracy level, and therefore can correctly fulfil EFSA needs and technical requirements in relations to the use case. There are many available solutions related to this use case that can be evaluated to choose the best one in terms of quality and costs. The cost of buying an already commercialised tool that needs to be customised based on EFSA needs and requirements is less than the amount spent on investing in research and development for a tool/service that does not exist yet, or that is in its initial level of technology readiness. Opportunities - Considering that the majority of the commercialised AI tools are developed for a general purpose, EFSA may need a unique solution tailored to its needs and context. However, most of the vendors have ensured their solution is highly customisable, even though it may require additional costs and time for implementation. A well-defined ontology is needed, which is currently not in place within EFSA. An ontology can help in defining the similarities between terms. Threats - The identification of the most relevant keyword within the literature through an AI-based solution would represent a good opportunity for EFSA to develop a well-defined domain specific ontology. The expected impacts of creating an ontology can be found within the horizontal recommendations section (5.2.1). - The adoption of AI technology would automate the identification of the most appropriate keywords for the search strategy, leading to time/efficiency gains and quality improvement. - One of the major issues of the PERA project is the resourceintensiveness when retrieving and analysing relevant information from databases and websites. SLR Keyword Identification can introduce time-efficiencies. There is a potential risk of incurring unexpected costs if the compatibility of the AI solutions within EFSA’s ecosystem is not properly assessed and ensured. Swot Analysis – SLR Relevant Abstract Screening Strengths - Weaknesses EFSA already has experience in the use of the current tool, DistillerSR, for Abstract Screening, Thus, there will be no significant resistance to change. www.efsa.europa.eu/publications 81 - During the consultation with EFSA staff, it emerged that many employees may be resistant to use a different tool for Abstract Screening, since DistillerSR is widely and satisfactorily used. EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence - - There are many available solutions related to this use case, which EFSA can evaluate through an effective and efficient procurement process, to eventually choose the best solution in terms of quality and costs. The cost of buying a commercialised tool that need to be customised based on EFSA needs and requirements is less than the amount spent on investing in research and development for a tool/service that does not exist yet, or that is in its initial level of technology readiness. Opportunities - Considering that the majority of the commercialised AI tools are developed for a general purpose, EFSA may need a unique solution tailored to its needs and context. However, most of the vendors have ensured their solution is highly customisable, even though it may require additional costs and time for implementation. - Domain-specific ontologies would need to be developed to enable the correct functioning of this solution. Threats - For testing purposes, EFSA may consider purchasing an additional tool for a trial period to compare its performance with that of DistillerSR. - The adoption of AI technology would automate the screening of papers’ abstracts and their analysis against the research question, leading to time/efficiency gains and quality improvement. - One of the major issues of the PERA project is the resourceintensiveness when retrieving and analysing relevant information from databases and websites. SLR abstracts screening can introduce timeefficiencies. In case EFSA decided to deploy an additional tool for testing and comparison with the existing tool (DistillerSR), there is a risk of users’ resistance. As a result, the new tool may be dropped, regardless of its effective performance, with all the effects that may be associated in terms of cost and time. Swot Analysis – SLR Literature De-duplication Strengths - - - Weaknesses Most of the tools identified do not separate the activities of data clustering and de-duplication. Therefore, in these cases, EFSA would purchase one solution that fulfils both the use cases. There are many available solutions related to this use case, which EFSA can evaluate through an effective and efficient procurement process, to eventually choose the best solution in terms of quality and costs. The cost of buying an already commercialised tool that needs to be customised based on EFSA needs and requirements is less than the amount spent on investing in research and development for a www.efsa.europa.eu/publications 82 - Considering that the majority of the commercialised AI tools are developed for a general purpose, EFSA may need a unique solution tailored to its needs and context. However, most of the vendors have ensured their solution is highly customisable, even though it may require additional costs and time for implementation. - When performing data de-duplication, it is important to not set too strict rules since this could lead to the risk of eliminating and losing relevant data. EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence tool/service that does not exist yet, or that is in its initial level of technology readiness. Opportunities - - - Without a complete ontology, this use case would only be used as a rule-based algorithm. Threats The adoption of AI technology would automate and make trustworthy the detection of duplicated papers highlighted for removal from SLR, leading to time/efficiency gains and quality improvement. - One of the major issues of the PERA project is the resourceintensiveness when retrieving and analysing relevant information from databases and websites. SLR Literature de-duplication can introduce time-efficiencies. Risk of incurring unexpected costs in case the compatibility of the AI solutions within EFSA’s ecosystem is not properly assessed and ensured. If rule-based tools (already used within EFSA) perform to a satisfactory level, it might be difficult to convince management to invest in an AIbased solution. Swot Analysis – Text Summarisation Strengths - - - - Weaknesses Most of the tools identified for this category, such as the IRIS AI, Risklick AI or IBM Watson Discovery, offer additional services/functions that can be used for other use cases. There are many available solutions related to this use case, which EFSA can evaluate through an effective and efficient procurement process, to eventually choose the best solution in terms of quality and costs. The cost of buying an already commercialised tool that needs to be customised based on EFSA needs and requirements is less than the amount spent on investing in research and development for a tool/service that does not exist yet, or that is in its initial level of technology readiness. - As most of the AI tools identified for this use case offer multiple services/functions, they may end up performing worse in comparison to solutions specialised in just a single use case. Although the implementation of the solutions may require some customisation, there are low risks of project failure, since the solutions already cover the main needs well. Opportunities www.efsa.europa.eu/publications Threats 83 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence - Most of the tools identified for this category, such as the IRIS AI, Risklick AI or IBM Watson Discovery, offer an integrated set of services/functions that can be used for other use cases as well, improving the overall user experience. - The adoption of AI technology would automate and simplify the summarisation of scientific evidence, leading to time and effort efficiency, quality improvement and reduction of human bias. The adoption of AI for this use case would be very relevant across EFSA, and not only for evidence management in risk assessment. - - There is a potential risk of incurring unexpected costs if the compatibility of the AI solutions within EFSA’s ecosystem is not properly assessed and ensured. One of the major issues of the PERA project is the resourceintensiveness when retrieving and analysing relevant information from databases and websites. Text Summarisation can introduce timeefficiencies. An additional pain point of PERA is the variability between the outcomes of peer-reviews conducted by different risk assessors. AI can help in this case through the application of topic modelling approaches or Automatic Text Summarisation. These can simplify the procedure and support PERA in producing informative summaries per each peer-review. These summaries might contain key elements of the review, which can be used in an AI tool (such as Comparable Appraisal Detection) to identify similarities and discrepancies. Swot Analysis – SLR Literature Findings Clustering Strengths - - Weaknesses This use case reports the largest number of identified commercialised tools in comparison with all the other prioritised use cases. Therefore, EFSA may choose the one which best fit its needs from many relevant solutions. Most of the tools identified do not separate the activities of data clustering and de-duplication. Therefore, in these cases, EFSA would purchase one solution that fulfils two use cases. www.efsa.europa.eu/publications 84 - Considering that the majority of the commercialised AI tools are developed for a general purpose, EFSA may need a unique solution tailored to its needs and context. However, most of the vendors have ensured their solution is highly customisable, even though it may require additional costs and time for implementation. - A complete subject-specific ontology is needed for this solution to give satisfactory results. EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence - The cost of buying an already commercialised tool that needs to be customised based on EFSA needs and requirements is less than the amount spent on investing in research and development for a tool/service that does not exist yet, or that is in its initial level of technology readiness. Opportunities Threats - Currently, in EFSA, data clustering is performed for short texts via the use of unsupervised learning techniques. Ideally, this should be improved to capture the full text. Therefore, a potential collaboration between EFSA (via providing data from its rich database) and a vendor may occur in the future. - The adoption of AI technology would automate and simplify the clustering of SLR findings in terms of relevance, leading to time and effort efficiency and quality improvement. The algorithm developed for the semantic analysis and reasoning, required for the clustering of SLR, may perform worse than a simpler AIenabled application - Vendor Management Scenario 2 – Innovation Procurement (PPI) Swot Analysis – Quality Improvement of the Scientific Output Strengths Weaknesses - The development of a solution aimed at improving the quality of scientific output might be considered a great investment, since it might also automatise other activities related to evidence management of risk assessment (e.g., it might be used to summarise the literature review, or it can syntheses evidence collected from other sources). - Although the implementation of the solutions will require some customisation, there are low risks of project failure, since the solutions already cover the needs well. Opportunities - The customisation of the solution might be quite challenging for EFSA considering the different features and functions that the solution could support to improve the quality of the scientific output (e.g., AI might support the grammar, synthesis, or syntax of the scientific output). - Launching an innovation procurement for this use case might result in a time-consuming and expensive process because of the need to perform a Proof of Concept/Pilot Project to demonstrate the solution’s performance and feasibility. Threats - www.efsa.europa.eu/publications 85 Even though a vendor may agree on customising its product to satisfy EFSA needs, there is no guarantee that the end-product will produce EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence - The adoption of AI technology would automate the proofreading and qualitative assessment of reports, leading to time and effort efficiency and quality improvement. accurate results. As a result, an amount of time and money could be invested without returning the expected result. Vendor Management Scenario 3 – Grant, PCP or Contest Swot Analysis – Expert pool identification Strengths Weaknesses - The development of an AI solution from scratch, or in its initial level of readiness through a grant or PCP might help EFSA in reducing overheads by building a tailored and customised solution. - The time required to build the proposed solution can extend far beyond the desired outcome. The challenges in training, to reduce false positives and increase quality production, may take years to overcome. - Even though the majority of the AI tools identified for this use case do not match exactly its purpose, they can still be used as a point of reference when developing a new tool. - To effectively develop this solution, there may be the need to integrate different components, such as web-crawling. - There is the risk of selecting and awarding the grant to the wrong candidate. This may result in investing a huge amount of time and money without delivering the expected solution. If experts’ information is not publicly available, the solution will not be able to deliver the desired results. Opportunities - Threats This solution might be applied in different units within EFSA. The adoption of AI technology would automate the identification of a list of relevant experts, leading to time and effort efficiency, enhanced trustworthiness, quality improvement, and the reduction of human bias. - The internal technical staff might not be prepared – from a technical point of view - to develop such as sophisticated AI solution. - Beyond the AI utility, before building web-crawling tools, there are a number of cybersecurity and GDPR issues to be taken into consideration. Swot Analysis – Expert selection Strengths Weaknesses - www.efsa.europa.eu/publications 86 The time required for the development of the solution can extend far beyond the desired outcome. The challenges in training, to reduce false EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence - - The development of an AI solution from scratch, or in its initial level of readiness through a grant or PCP might help EFSA in reducing overheads by building a tailored and customised solution. Developing an AI solution from scratch would enable EFSA to own the intellectual property of the AI solution in this specific area of evidence management. Opportunities - - positives and to increase quality production, may take years to overcome. There is the risk of selecting and awarding the grant to the wrong candidate. This may result in investing a huge amount of time and money without delivering the expected solution. Threats The adoption of AI technology would automate the ranking and selection of relevant experts, leading to time and effort efficiency, enhanced trustworthiness, quality improvement and reduction of human bias. - No threats were identified for this use case. Swot Analysis – Comparable appraisal detection Strengths Weaknesses - The development of an AI solution from scratch, or in its initial level of readiness through a grant or PCP might help EFSA in reducing overheads by building a tailored and customised solution. - Developing an AI solution from scratch would enable EFSA to own the intellectual property of the AI solution in this specific area of evidence management. - - - Opportunities Threats - A research project in this area is likely to contribute important findings to the wider NLP scientific community. - The adoption of AI technology would automate the appraisal of evidence gathered from comparable research questions, leading to www.efsa.europa.eu/publications A specific functionality, which includes semantic analysis and reasoning, for automating the use case might be difficult to develop as it will require a lot of research activity, given that it is not adequately developed. Building and training an algorithm while reducing false positives and ensuring quality production may be time and effort consuming (taking years). By launching a grant or a PCP for the development of a solution, EFSA might invest a huge amount of money without receiving the expected solution. 87 - The internal technical staff might not be prepared – from a technical point of view – for developing and maintaining the AI solution. EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence time and effort efficiency, enhanced trustworthiness, quality improvement and reduction of human bias. www.efsa.europa.eu/publications 88 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence 4.3. Prioritisation of Working Areas In the previous sections, a list of horizontal and vertical recommendations was provided with the aim to: 1) Build EFSA’s readiness to adopt AI for evidence management in risk assessment (and in a wider sense) and 2) Enable the adoption of AI in the specific prioritised uses via the three proposed AI adoption scenarios. The two points together will allow EFSA to reach its vision to increase in the accessibility and breadth of the body of evidence, thereby enhancing the trustworthiness in the risk assessment process, and to apply human centric artificial intelligence in close co-existence with human expertise by 2027. However, to concretely support EFSA in its decision making and resources allocation, we present in this section a prioritisation of these vertical and horizontal recommendations to enable their order and dependencies to be understood. To prioritise the horizontal recommendations, we have considered the following elements: • The impact on / need within evidence management of risk assessment (i.e., in terms of it being a prerequisite for the adoption of AI, or a “best-practice” recommendation), • The impact of the recommendation across EFSA (i.e., not only in the scope of evidence management in risk assessment, but also in other key areas within EFSA, based on the information and needs collected from internal stakeholders, and from other roadmaps. For example, input from the NAMs team has been taken into consideration, including their need for automating the area of data standardisation and integration, which might be covered by the infrastructure and governance recommendations.) • The dependence of each recommendation on other horizontal recommendations. It is worth noting that the table below (Table 18) provides an account of the effort required to adopt each vertical and horizontal recommendation. Regarding the horizontal recommendations, which are foundational to the development of EFSA’s AI capability, the effort is not estimated, as the projects need to take place regardless of the effort. We then proceeded with incorporating in the prioritisation the vertical recommendations for the different use cases, on the basis of: • The dependence on horizontal recommendations, and • The anticipated impact on evidence management in risk assessment (as per the impact assessment set out in Table 17). Finally, the two prioritisation exercises (for horizontal and vertical recommendations) were integrated to develop a consistent and comprehensive roadmap aimed at delivering to EFSA quick wins, together with a sustainable and scalable AI capability development. For example, although the creation of domain specific ontologies (Horizontal Recommendation 1) is not a priority compared to other more foundational capacity development actions, the rapid adoption of some market-ready solutions relies on the existence of domain specific ontologies. As a result, the development of EFSA-specific ontologies was prioritised to consider this relationship. In addition, domain specific ontologies will also support the scientific advancement of AI in those related areas. Rationale behind ratings For the horizontal recommendations, the rationale behind the ratings is the following: • Impact/Need within Evidence Management: Rated as “High” if it is a pre-requisite, and “Medium” if it is a best practice for the adoption of AI; • Effort: rated in terms of the time and financial investment needed to adopt the recommendation; • Impact across EFSA: rated on the basis of the needs expressed by other EFSA stakeholders consulted and roadmaps (e.g., NAMs and PERA roadmap) in terms of the impact that this recommendation would have on www.efsa.europa.eu/publications 89 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence their own work. Indeed, some horizontal recommendations (e.g., Implementation of a Data Infrastructure) are big projects that would be beneficial across other areas in EFSA, such as for the NAMs team, in order to deal with data-intensive activities. For the vertical recommendations, the rationale behind the ratings is the following: • Impact/Need within Evidence Management: Impact ratings based on the previous analysis. • Effort: rated as “Low” for use cases falling under Scenario 1, “Medium” for those falling under Scenario 2, and “High” for those falling under Scenario 3. • Impact across EFSA: rated on the basis of the needs expressed by other EFSA stakeholders consulted and roadmaps (e.g., NAMs team and PERA roadmap) in terms of the impact that this recommendation would have on their own work. First, the overall prioritisation was performed based on the recommendations’ dependencies. Then, in terms of the impact ranking (both within evidence management and across EFSA) and, lastly, on the basis of the effort needed (the latter only in the case of the vertical recommendations). Table 17 - Prioritisation of Horizontal and Vertical Recommendation No 1 2 3 4 5 6 No UC 1 UC 2 UC 3 UC 4 Horizontal Recommendation Create (or adopt existing) ontologies for domains of relevance for EFSA 57 Implement a Data Infrastructure which enables Big Data and AI development Design and Implement a Data Governance framework which addresses AI Adopt DevOps Adopt a Trustworthy AI framework 58 Develop Data Science Competencies in NLP and design behaviour change mechanisms Vertical Recommendation Data Collection Terminology Assessment SLR Keywords Identification SLR Relevant Abstract Screening SLR Literature Deduplication Capability Development Dimension Impact/Need within Evidence Management Effort Impact across EFSA Dependencies Data & Information High Medium Low 6 Technology & Infrastructure High High High 3 Organisation & Governance High High High None Medium Low Medium 2,3,6 Medium Medium High None Skills & Culture High High Low None Capability Development Dimension Impact Within Evidence Management Effort Impact Across EFSA Dependencies with horizontal recommendatio ns Use Cases 2.25 Low Medium None Use Cases 2.5 Low Low (1)59,6 Use Cases 1.75 Low Low 6,1 Use Cases 2 Low Low 1,6 Process & Integration Process & Integration 57If the ontology is to be developed in-house, it would have a dependency with horizontal recommendation 2, Implement a Data Infrastructure which enables Big Data and AI development 58 An alignment with HR 3 is recommended as best practice. 59 It is suggested to have an ontology, although this is not mandatory as most of the solutions for this use case do not require it. www.efsa.europa.eu/publications 90 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence UC 5 UC 6 UC 7 UC 8 UC 9 UC 10 Text Summarisation SLR Literature Findings Clustering Quality Improvement of the Scientific Output Expert Pool Identification Expert Selection Comparable Appraisals Detection Use Cases 2.25 Low Medium (1)60, 6 Use Cases 2.5 Low Medium 1, 3, 2, 6 Use Cases 2 Medium Low 1,2,3, 6 Use Cases Use Cases 2.75 3 High High Low Low 2,3,6 2,3,6 Use Cases 2.5 High Medium 1,2,3,6 Based on the above assessment, all of our recommendations (both horizontal and vertical) have been prioritised and ordered as shown below in Table 18, along with the estimated time-frames. The approach used to distribute the projects along the time-frame followed the dependencies and urgency of each project, as mentioned earlier, but also in an attempt to avoid placing too many projects in the same time period, in order not to overcharge EFSA's effort. The table allows the quick-wins that EFSA can obtain to be visualised, while developing its overall AI capability (horizontal recommendations). This approach is fundamental to achieve EFSA’s 2027 vision 61. The proposed time-frames do not take into account the time to prepare, launch and award the procurements/grants. In addition, it should be noted that, in the event that horizontal recommendations become projects to be carried out in house (and not outsourced via procurement), a 30%-time increase must be foreseen. 60 it. 61 It is suggested to have an ontology, although this is not mandatory as most of the solutions for this use case do not require EFSA, Theme (concept) paper on Artificial Intelligence in risk assessment. www.efsa.europa.eu/publications 91 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Table 18 - Roadmap for actions (Horizontal & Vertical Recommendations) Recommendation H2 2022 H1 2023 H2 2023 H1 2024 H2 2024 Programme management H1 2025 H2 2025 H1 2026 H2 2026 H1 2027 H2 2027 Programme management Develop (or adopt existing) Ontologies relevant to EFSA Horizontal 1 Continuous maintenace required for this recommendation Adopt a Data Infrastructure for Big Data and AI Horizontal 2 Horizontal 3 Design and Implement Data Governance Training required for this recommendation Horizontal 4 Implement DevOps Training and adoption Adopt a Trustworthy AI framework - Risk M gmt Framework Adopt a Trustworthy AI framework - CoC Horizontal 5 Training required for this recommendation Training required for this recommendation First round of recruitment Assessment of Data Science Skills and recruitment/upskilli ng plan Horizontal 6 Second round of recruitment* Upskilling of current staff Scenario 1 UC 1 Scenario 1 UC 2 Scenario 1 UC 3 Vertical scenario 1 Scenario 1 UC 4 Scenario 1 UC 5 Scenario 1 UC 6 Scenario 2 UC 7 Vertical scenario 2 Scenario 3 UC 8 Vertical scenario 3 Scenario 3 UC 9 Scenario 3 UC 10 www.efsa.europa.eu/publications 92 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Since there are multiple projects relating to the adoption of AI for evidence management, it is appropriate to also consider some additional effort related to programme management.. The programme management may have the following objectives: • • • • • • • Ensuring that AI projects are aligned with the overall EFSA strategy; Planning, coordinating, and monitoring of progress, and managing of all the AI projects’ inter-dependencies, including oversight of any risks and issues arising; Facilitating EFSA governance by bringing together the internal key stakeholders from different units impacted by the change e.g., through the organisation of bi-monthly progress meetings; Providing strategic advice on the design of the communication and engagement plans with the interested stakeholders (see chapter 6); Managing third party contributions to the programme e.g., by organising an internal EFSA forum to connect and align all involved project managers and contractors; Ensuring that there is an efficient allocation of resources and skills within the programme's individual projects; Enabling effective change and endorsement of all AI projects by EFSA staff, by supporting the definition of ad-hoc change management actions (e.g., design of a communication plan, training plan, resistance management plan). Budget and Time Estimates This section aims to provide some key details for each of the six horizontal recommendations and for the three vendor scenarios (vertical recommendations), such as their estimated budget, time and effort (expressed in man days). Initially, some general assumptions are presented, which relate to the estimates for all of the projects. Then we show the estimated figures, together with some more specific assumptions that consider the peculiarities of each horizontal or vertical project. General Assumptions As mentioned before, the following assumptions are valid for all projects: • • • • • The projects’ activities can be performed in-house or outsourced to contractors; Standard market rates with no discounts were considered (500 EUR/Day for junior positions, 750 EUR/Day for middle management positions, 1000 EUR/Day for senior positions). These costs should be adjusted with EFSA rates, in case EFSA decides to undertake the projects in house. Timings were estimated considering an outsourcing scenario. Alternatively, in an in-sourcing scenario, we suggest counting an additional 30% of time required for implementing the projects, due to the current limited availability of internal staff. Procurement costs were not considered (i.e. for the definition of the proposal, selection and awarding, and administrative costs). For the implementation of the vertical recommendations, costs and times for the scouting activity and vendor selection were not considered. In the future, if EFSA decides to apply the same approach to other use cases (especially for those outside of evidence management), it should also count these cost and time factors for scouting and vendor selection in order to obtain a complete estimate. Estimates and specific assumptions www.efsa.europa.eu/publications 93 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Programme Management The table below sets out the expected budget, time scale, man days and specific assumptions for each horizontal recommendation. Table 19 – Expected budget and time scale of the Horizontal Recommendations. Budget 1) Develop (or adopt existing) ontologies for domains of relevance for EFSA 483K € 2) Implement a Data Infrastructure for Big Data Analytics & AI development Proposed time scale 1 year project + 4 years maintenance Man Days (md) Specific Assumptions 708 Days • Estimated costs for the development of 3 ontologies • Considered 20k EUR/Year*5 years of licence costs, to be deleted in case of open-source tools • Considered 200K to develop the ontologies in case EFSA starts from scratch (to be reduced by 30% if you start with available ontologies) 764 Days • 150K per year were estimated for the infrastructure (Azure, *50 users, * 100 TB, pay as you go) • The logical and physical data modelling was estimated for 10 subject areas 676 days • The licence cost for the tool is not included (if not already available, the licence for data quality + stewardship/ownership + Master Data Management to be purchased) 608K € 1 year project + 1 year training/upskilling 417K € 1 year project + 0.5 year training/upskilling 4) Establish Development & Operations Management (DevOps) 366K € 0.5 year project, including training 504 days • The licence cost for the infrastructure/tool is not included (it is assumed to be included in the Azure bundle) 5) Adopt a Trustworthy AI Framework 394.5K € 4 months for the CoC + 1 year RMF + 4.5 year training 630 days NA 0.5 years + 5 years training/upskilling of current staff 270 days NA 3) Implement a Data Governance framework which addresses AI 6) Expand data science competencies & design behaviour change mechanisms 183K € Total Estimated Budget 2022-2027 2.4Mln € As regards the vertical recommendations – including scenarios 1,2,3 - their estimated budget and time is reported in Table 20 below. Table 20 - Expected budget for Vertical Recommendations. Budget per Use Case Vertical Recommendation Scenario 1 58 K € (including 1 year tool licence) www.efsa.europa.eu/publications Time No of cases 4 months 6 94 use Use Case Titles Data Collection Terminology Assessment SLR Keywords Identification SLR Relevant Abstract Screening SLR Literature De-duplication Text Summarisation SLR Literature Findings Clustering Budget * No of use cases 348K € EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Vertical Recommendation Scenario 2 215 K € (including 1 year tool licence) Vertical Recommendation Scenario 3 367 K € 8 months + 1 year maintenance activities 18 months +1 year maintenance activities 1 Quality Improvement of the Scientific Output 215K € 3 Expert Pool Identification Expert Selection Comparable Appraisals Detection 1.1Mln € Total Estimated Budget 2022-2027 1.7Mln € The following table sets out the estimates for the programme management. Table 21 - Expected budget, time scale and man days for Programme Management. Budget Programme Management Proposed time scale 462K € Man Days (md) 5.5 years 660 days The following table summarises the estimated overall costs from 2022 to 2027. Table 22 - Expected Total Budget per Horizontal and Vertical Recommendations. Project Type Budget Programme Management 0.46Mln € Horizontal Recommendations 2.4Mln € Vertical Recommendations (Scenarios * 10 use cases) 1.7Mln € Total Estimated Budget 2022-2027 4.56Mln € www.efsa.europa.eu/publications 95 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence 5. Communication and Engagement Plans This chapter aims to identify potential communication and engagement opportunities in the area of AI approaches for evidence management of the risk assessment, and to outline the connected actions, taking into consideration both EFSA’s objectives and potential stakeholders’ needs 62. The main outputs are a communication and commitment plan, respectively designed to make EFSA's stakeholders aware about the adoption of AI in the evidence management process, and to take on board key potential partners to support the development and implementation of AI-based solutions. The chapter is structured as follows: • • • • • 5.1. Section 5.1 presents the general approach to build a communication and engagement plan in view of the AI adoption for evidence management. First, the approach consists of the mapping of EFSA stakeholders’ categories, their assessment through a multi-criteria methodology and the design of the communication and engagement plans Then, focus on the results of the stakeholder's categories mapping is provided (5.1.1). Finally, the criteria used for the stakeholder's assessment, together with the results, are described (5.1.2). The results of the assessment are represented through a two-axis matrix with four quadrants, each associated to a specific result, where stakeholders are placed. Section 5.2 focuses on the lower-left quadrant of the matrix, which includes the stakeholders that are not significantly valuable for the purpose of engagement activities, and that are not particularly sensitive to the AI topic. In this scenario, the optimal strategy would be monitoring these stakeholders’ future actions around AI and, in parallel, keeping them updated on EFSA's AI work. Section 5.3 focuses on the upper-left quadrant of the matrix, where stakeholders are still not particularly relevant for the purpose of engagement activities in view of AI adoption, but may be particularly sensitive to the AI topic with specific concerns on the matter. For these organisations, the design of an ad-hoc communication could be the optimal strategy to respond to their concerns. Section 5.4 outlines the proposed engagement plan for those stakeholders’ categories that are particularly relevant for the purpose of engagement activities in the field of AI adoption in the evidence management process (lower-right and upper-right quadrants of the matrix). Section 5.5 puts forwards some final considerations around the communication and engagement plans, regarding their time for implementation, dependencies with other horizontal recommendations, and scope. General Approach Not all of the stakeholder groups are of equal importance to a public organisation. Knowing when, how, to whom and why to send communications and/or collaborate is vital to any communication/engagement plan. Some stakeholders may require immediate notification about a change or situation affecting the organisation, whilst others that are less impacted or not involved at all do not need access to the same information. To lay down a proper communication and engagement plan, we designed the following three-step approach: 1. Mapping of Stakeholders: consists of the mapping of EFSA’s relevant stakeholders to be targeted, identifying them according to specific categories (e.g., Industry, Academia) and sub-categories (e.g., NGOs, Business Operators). As mentioned before, we refer to both stakeholders within EFSA’s food safety ecosystem for which EFSA is accountable, and potential partners that may collaborate with EFSA for the purpose of this project around the AI adoption. 2. Assessment of Stakeholder’s Categories: consists of assessing each stakeholder category/sub-category based on several criteria that are designed for the purpose of this project. The aim of this assessment is to We refer to the different stakeholders’ types mapped through the previous project’s activities to be potentially engaged for the purpose of the project (see sections 3.1 and 3.3), and also other actors within EFSA’s food safety ecosystem to be informed about the AI developments (e.g. EFSA’s risk managers, NGOs etc.). More details are reported in the following sections. 62 www.efsa.europa.eu/publications 96 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence understand how better to approach and/or involve the different stakeholders in view of the implementation of AI technology for evidence management. 3. Development of the Communication Plan and Engagement Plan based on the assessment of stakeholder’s categories/sub-categories, including specific actions to be followed and methods to be to be used. Specifically, stakeholders that are not relevant for collaboration opportunities, but that should be kept informed on EFSA’s AI developments are targeted for the communication plan, while key stakeholders that can contribute to the AI adoption in EFSA will be targeted for the engagement plan. Further details on the methodology and results of the above-mentioned steps are provided in the following sections. Figure 32 - General Approach for Communication and Engagement. 5.1.1. Mapping of Stakeholders’ Categories/Sub-categories As explained previously in the general approach, the first step consisted of the mapping of EFSA stakeholders, clustering them according to specific categories (e.g., Industry, Academia, and sub-categories (e.g., NGOs, Business Operators, Distributors, etc). Table 23 provides an overview of EFSA’s stakeholder categories and sub-categories, developed by adapting and extending the official categorisation of EFSA food chain stakeholders - adopted by the Management Board in 2016 63 and reflected in EFSA’s Stakeholder Engagement Approach – to the scope of this project around AI adoption for evidence management. In particular, our categorisation also includes additional stakeholders identified during the project’s data collection activities, which are relevant for the purpose of collaboration to foster AI adoption in EFSA. Table 23 - List of EFSA stakeholders' categories and sub-categories. Stakeholder Category Stakeholder Sub-category European Commission Sister Agencies Institutions Other European Agencies Other international organisations National Competent Authorities (incl. Art. 36 org) Industry 63 Business and food industry IT/Software companies https://www.efsa.europa.eu/en/partnersnetworks/stakeholder www.efsa.europa.eu/publications 97 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence University spin-offs AI start-ups Distributors and HORECA [Hotels, Restaurants, Catering] Farmers and primary producers Advocacy groups NGOs and civil society organisations Consumer organisations Practitioners’ associations Academia Universities and research centres (incl. Art. 36 org) Digital Innovation Hubs Development Levers64 Technology transfer offices Incubators and accelerators 5.1.2. Assessment of Stakeholder’s Categories Once the stakeholder categories/sub-categories were identified, each of them was assessed based on three criteria to understand how to involve and approach them in view of the adoption of AI for evidence management. • Relevance for the project, expressed as the strategic importance that stakeholders have in terms of concrete cooperation/collaboration opportunities with EFSA in view of AI adoption for evidence management. This reflection is based on the analysis performed in section 3.3, where the stakeholders playing a prominent role are mapped throughout the phases of AI development lifecycle to fulfil EFSA's different needs (from 0 to 5). • Impact in terms of effect – positive or negative – on EFSA’s accountability of stakeholders’ opinions with respect to the implementation of AI projects, or in terms of effective contribution of their solutions or actions to enable EFSA’s adoption of AI technology. Therefore, this criterion shows the importance of having a stakeholder on board by establishing pertinent collaboration/cooperation models or properly communicating with them on EFSA’s work and progress. • Easiness to collaborate with the stakeholders, based on current or past collaboration and the presence of existing mechanisms and processes that facilitate collaboration in AI-related projects. In other words, the easiness to collaborate depends on the kind of relationship / level of interaction of EFSA with a given stakeholder, which can be categorised as established, evolving and emerging. For example, stakeholders with an established relationship with EFSA are those that are that have an active collaboration/exchange with EFSA; stakeholders with an evolving relationship are those that have not extensively collaborated with EFSA in the past and are willing to strengthen the relationship, and stakeholders with an emerging relationship levels are those organisations that have never collaborated with EFSA with respect to AI development. Finally, stakeholders that are not able/interested to contribute are classified as Not applicable (e.g., NGOs usually do not have the proper expertise to collaborate with EFSA in this field). In fact, this criterion is only applied to stakeholders that can contribute to some extent to the adoption of AI solutions, whether by commercialising them (e.g., software companies and startups), providing access to expertise (e.g., DIHs, Universities, Accelerators, etc), exchanging best practices and lessons learned, or contributing with EFSA by co-developing AI solutions (e.g., EU and National Institutions), etc, as mentioned in section 3.3. 64 Development levers are entities or programmes that accelerate the modalities to engage the AI community, and establish cooperation. Therefore, they can connect EFSA with other stakeholders involved in innovation processes. www.efsa.europa.eu/publications 98 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence It is important to highlight that the assessment was based on the information collected through the various methods during the data collection and analysis phases of the project. Based on the results of the assessment, stakeholder sub-categories were positioned in a matrix, as shown in the figure 34, where the first criterion is represented in the horizontal axis, the second in the vertical axis, and the third is represented through a colour code. Figure 33 - Communication/engagement strategies (Impact/Relevance/Easiness to collaborate matrix). for the identified stakeholder segments Once the stakeholders’ categories were positioned, the matrix was split into four quadrants, each grouping the categories sharing similar results and the same outcome in terms of communication or engagement actions: • • • [LOWER-LEFT] Inform and Monitor/Listen: includes stakeholders associated to a low relevance and low impact, such as farmers and producers, distributors and HORECA, and the business and food industry. These stakeholders are not significantly valuable for the purpose of engagement activities in the field of AI adoption in the evidence management process (i.e., they do not offer any AI solution, do not undertake any relevant initiative around AI development, or have any solid AI knowledge). In addition, as they are not sensitive to the AI topic, no concerns can be expected from their side in case EFSA launches a public communication around the development of AI in the organisation. In this scenario, the optimal strategy would be monitoring these stakeholders’ future actions around AI and, in parallel, keeping them informed on EFSA’s progress on AI adoption through the available institutional means and channels (e.g., press release, EFSA website), without the necessity to build personalised communications messages. [UPPER-LEFT] Design an ad-hoc communication: includes stakeholders associated to a low relevance and high impact, such as practitioners associations, advocacy groups, and consumer organisations. Indeed, while these stakeholders are not significantly valuable for the purpose of possible collaboration activities in view of AI adoption for evidence management, they are particularly sensitive to the AI topic. This last aspect may entail possible concerns and claims from their side, in case of a public communication of EFSA’s activities in the field of AI, with a potential negative impact on the organisation’s reputation. Hence, it would be recommended to design an ad-hoc communication plan that considers their concerns on AI, and mitigates the risk of potential claims and legal actions. [LOWER-RIGHT] Engage in the ecosystem: includes stakeholders associated to a high relevance, low impact organisations that have never collaborated with EFSA with respect to AI development (EFSA’s AI ecosystem), such as international organisations (e.g. FAO), university spin-offs, and the so-called development levers (DIH, Accelerators, and Technology Transfer Offices). These stakeholders are www.efsa.europa.eu/publications 99 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence relevant for the purpose of engagement/collaboration activities in the field of AI adoption for evidence management (e.g., development levers can accelerate the modalities to establish cooperation with other actors, such as AI start-ups and universities with relevant AI expertise). On the other hand, these stakeholders are considered less impactful - compared to those within the up-right quadrant - since they do not specifically offer any AI solution at the moment, or otherwise given their mission. For instance, on the one hand, development levers do not market or develop AI solutions, but do have the function of accelerating cooperation with other stakeholders. On the other hand, with regard to International organisations (e.g., FAO) and university spin-offs, according to our research they do not currently commercialise any AI solutions for evidence management, or do not pursue any initiative/action around AI adoption in such area. In the light of this, the optimal strategy would be engaging these stakeholders within EFSA's AI ecosystem (intended as those organisations that have already collaborated with EFSA in AI development) to create synergies, open up opportunities for future concrete collaboration/cooperation, and boost the innovation process. • [UPPER-RIGHT] Maintain & Reinforce Collaboration: includes stakeholders with a high relevance, high impact and that are already part of EFSA's AI ecosystem, such as Universities and research centres, Software companies, EFSA’s sister agencies and other EU agencies, Article 36 organisations, European Commission bodies and National Authorities. These stakeholders are particularly relevant for the purpose of engagement activities in the field of AI adoption in the evidence management (e.g., software companies offer a high number of commercialised AI solutions; some EU Agencies are developing AI solutions that fall within the area of evidence management and are open to collaborate in the future with EFSA), and having them on board would generate a positive impact for EFSA. In the light of this, the optimal strategy would be maintaining and reinforcing the collaboration actions with these partners in the future for the adoption of AI. Figure 34 - Communication/Engagement strategies for the identified stakeholders segments. 5.2. Inform & Monitor/Listen Matrix Quadrant As anticipated in the previous section, EFSA stakeholders with a low relevance and impact were positioned on the lower-left portion of the matrix and are mainly composed of: Farmers and primary producers, Business and food industry, Distributors and HORECA. These stakeholders are not relevant for the purpose of collaboration/cooperation activities with EFSA in the field of AI adoption for evidence management, and they do not have any particular concern related to the use of AI (with rare and isolated exceptions). This is proven by the fact www.efsa.europa.eu/publications 100 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence that no position papers or articles, published by them, were found on this topic. Therefore, no concerns could be expected from their side in the event of a public communication of EFSA’s activities in the field. Hence, these stakeholders should be monitored and kept informed on the AI technology adoption within the organisation through the available EFSA means and channels (e.g., press release, EFSA website), without the need to build an ad-hoc communication plan. It is important to keep in mind that, due to the emerging importance of AI, interests and concerns related to this topic may change for these stakeholders. Thus, if during the monitoring activities it emerged that these stakeholders have developed a higher regard and sensitivity towards the AI topic in the future, they would be moved to the upper quadrant on the left (Higher Impact), and therefore targeted for an ad hoc communication plan. Figure 35 - Inform & Monitor/Listen Quadrant. 5.3. Communication Plan As explained in section 5.1.2, Stakeholder sub-categories with a low relevance and high impact, falling into the upperleft quadrant of the matrix, were targeted for the development of an Ad-hoc Communication plan on the adoption and use of AI technology. Figure 36 - Ad-hoc Communication Quadrant. To develop a sound and effective Communication Plan, three main steps can be followed: www.efsa.europa.eu/publications 101 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence 1. Identification of key stakeholders’ concerns related to the adoption of AI. This activity has been carried out through a desk research to explore position papers, articles and reports published by EFSA’s stakeholders with an interest in the Authority’s work, or in the wider food and feed sector; 2. Definition of potential risks associated with the communication to the targeted stakeholders of the adoption of AI in the evidence management, based on their concerns. 3. Definition of mitigation actions for the risks identified, aimed to build proper communication messages – in line also with some key horizontal recommendations described in chapter 4 - around EFSA’s AI work. 5.3.1. Societal concerns on AI Adoption To identify the concerns of EFSA’s stakeholders in relation to the potential adoption of AI, an ad hoc desk research was conducted. During this task, position papers, articles and reports presenting stakeholders’ views on the adoption of AI were investigated, per each stakeholder’s category/sub-category. Although it emerged that different stakeholders share similar concerns, we managed to associate them to specific sub-categories. The main needs and concerns on AI adoption are related to the following domains: • • • • • • • • • • Ethics, trustworthiness and fairness in AI Liability, accountability and right to compensation, if AI practices harm the public in some way Transparency Environmental impact Bias and discrimination Privacy and data protection Manipulation Legal compliance Digitalisation Access to justice This section presents further details and some examples from our research. Various position papers were found for the NGOs and civil society organisations category, in particular for consumer organisations. The European Consumer Organisation (BEUC), for instance, has published a position paper that focuses on how Artificial Intelligence practices could harm the consumers. The paper65, called “Regulating AI to protect the consumer”, provides insights regarding concerns and risks that could be derived from AI applications. According to this, although AI technology could provide great improvements and make consumers’ life more convenient, it also comes with great risks and could have major implications for consumers’ autonomy and self-determination, their privacy, their capacity to interact with products and services and ultimately, in the ability to hold businesses responsible if something goes wrong. The paper refers to the proposed AI Act 66, as it should provide consumers with the rights and protections they need to be at ease when using AI, ensuring that the EU’s fundamental rights and values are respected. Though the organisation welcomes the attempt of the European Commission to regulate and provide a legal framework for Artificial Intelligence, BEUC believes that the proposal requires substantial improvements to guarantee that consumers have the protections they need, and can trust AI to respect their rights and freedoms. In general, when applying AI, some basic principles should be taken into consideration: 65 BEUC, Regulating AI to Protect the Consumer, (2021). Available at: https://www.beuc.eu/publications/beuc-x-2021088_regulating_ai_to_protect_the_consumer.pdf 66 European Commission, Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence and Amending Certain Union Legislative Acts, (2021) available at https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52021PC0206 www.efsa.europa.eu/publications 102 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence • • • Fairness Accountability Transparency In addition to these basic principles the paper provides a list of forbidden AI practices that could potentially harm consumer rights: - AI practices should not manipulate consumers in a way that could cause them economic harm; AI regulations should protect consumers made vulnerable through the use of persuasion profiles and personalisation practices (digital asymmetry); A ban on AI used by public and private organisations to evaluate individuals on the basis of their social behaviour and personal attributes, in order to avoid consumer discrimination. According to the paper, in order to be protected, consumers should have the right to be given clear explanations about how an AI system could affect their work, and should have the right to object to algorithmic decisions. In this sense, AI decisions should always be backed by human oversight, meaning that it should be a human-centric AI. In addition to the information provided by BEUC, according to our research, organisations such as Consumer International, ANEC, and ICPEN consider it relevant to protect consumers rights when applying AI to specific contexts. Moreover, other organisations that do not belong to the Consumer organisation category consider it important to address matters such as transparency, accountability, privacy and data protection, etc, for example advocacy groups and practitioners’ associations that work in the health care sector (AESGP, EPHA, FVE and EFA). These considerations are reinforced by those expressed by the report 67 of the European Union Agency for Fundamental Rights (FRA), which illustrates some of the ways that companies and the public sector in the EU are looking to use AI to support their work, and whether and how they are taking fundamental rights considerations into account. Even though the Agency is positioned in the higher part of the relevance matrix, we mention it here since its view is linked with those previously presented. The main findings of the FRA report can be summarised as follows: • • • • • Using AI systems can affect a wide range of fundamental rights, regardless of the field of application. These include privacy, data protection, non-discrimination and access to justice. So, when introducing new policies and adopting new legislation on AI, the EU legislator and the Member States, acting within the scope of EU law, must ensure that respect for the full spectrum of fundamental rights. Prior impact assessment mainly focuses on technical issues, and rarely address potential effects on fundamental rights, so the EU legislator should consider making mandatory impact assessment that covers the full spectrum of fundamental rights before any AI system is used. The EU and Member States should ensure that effective accountability systems are in place to monitor and, where needed, effectively address any negative impact of AI systems on fundamental rights. They should consider, in addition to fundamental rights impact assessments, introducing specific safeguards to ensure that the accountability regime is effective. EU Member States should consider encouraging companies and public administration to assess any potentially discriminatory outcomes when using AI systems. The European Commission and Member States should consider providing funding for targeted research on potentially discriminatory impacts of the use of AI and algorithms. To effectively contest decisions based on the use of AI, people need to know that AI is used, and how and where to complain. Organisations using AI need to be able to explain their AI system and decisions based on AI. The EU legislator and Member States should ensure effective access to justice for individuals in cases involving AI-based decisions. To ensure that available remedies are accessible in practice, the EU legislator and Member States could consider introducing a legal duty for public administration and private companies using AI systems to provide those seeking redress information about the operation of their AI systems. 67 FRA, Getting the Future Right: Artificial Intelligence and Fundamental Rights, (2020). Available at: https://fra.europa.eu/sites/default/files/fra_uploads/fra-2020-artificial-intelligence_en.pdf www.efsa.europa.eu/publications 103 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Other papers published by NGOs and civil society organisations tackle different aspects on how AI could impact their specific fields of competence. For example, the European Environmental Bureau has shared on its website a paper68 published by the European Parliament that focuses on the potential of AI to help to achieve a green transition, focusing on environmental potential, characteristics and causes of environmental risks that could be derived from applying AI to this specific field. For instance, AI could have negative impacts on the environment, since the use of digital, and hardware and infrastructure such as data centres and networks, could lead to an increased consumption of material and resources of energy. Other risks could derive from systemic effects such a rebound effect, that could evolve as a consequence of intended or unintended changes in the behaviour of consumers, users or producers. Such effects can be the consequence of opaque dynamics of AI system learning, running contrary to an intended environmentally friendly function of an application. For example, products designed to automatically manage and implement more efficient energy consumption may actually have the effect of causing users to give up control over their energy consumption and over-consume. Other adverse environmental impacts could result from AI applications that, as a consequence of their intended employment, contribute to GHG emissions and nature destruction, for example the use of AI to unlock oil and gas deposits, and to explore and develop new territories for fossil fuel extraction. In addition, the paper highlights that the proposed AI Act of the European Commission does not include any hazards related to the environment unless adverse environmental impacts pose a direct threat to human rights or interests. For instance, rules related to data governance, transparency, human oversight and security do not yet provide a governance system that will avoid adverse environmental impacts, so the EEB underlines the need for further research and development of methods and institutional procedures for assessing and controlling environmental risks caused by AI. Other organisations that deal with the environment such as Greenpeace have concerns similar to those of EEB. With regard to NGOs and civil society organisations that fall within the health and medical sector, such as AESGP69 and EPHA70, the position that emerged from their papers are overall aligned. The use of AI is seen as potentially beneficial, though it should only serve as a tool that could optimise day-to-day operations and data analysis, and it should never substitute nor erase in-person healthcare services. These organisations point out that supervision of AI should be proportionate to the intended use, and led by defined risk categories with clear internationally defined criteria, focusing on topics such as the data protection of the patients. In addition, as pointed in other papers and articles on the development and use of AI-powered tools, accountabilities when using AI in the healthcare sector need to be clearly defined. Also, AI should be Ethical and Trustworthy, meaning that AI systems need to be human-centric, with the aim of increasing individual and societal well-being. In conclusion, since the healthcare sector is a highly regulated environment, regulators must ensure a fair level playing field for all actors by developing and enforcing horizontal regulations across sectors. Furthermore, EuroCommerce, which is an organisation that aims to favour European retail, wholesale and trade, has published articles71 that overall consider AI as a powerful and useful tool that could have a huge potential for the development of the world economy. They consider AI as a tool that will be crucial to better understand customer needs, creating services that they want and making online and offline sales more efficient. In addition, they believe that: • • Having a positive narrative towards AI technologies is a prerequisite to unlock Europe’s tech sovereignty The future EU framework for AI should be technology-neutral and focus more on achieving desirable outcomes rather than regulating AI tools. 68European Parliament, The role of Artificial Intelligence in the European Green Deal,(2021) available at: https://www.europarl.europa.eu/RegData/etudes/STUD/2021/662906/IPOL_STU(2021)662906_EN.pdf 69 https://aesgp.eu/content/uploads/2021/08/AESGP-response-to-the-AI-regulation-proposal_FINAL.docx.pdf 70 https://epha.org/the-patients-voice-an-integral-part-of-digital-health-transformation/ 71 https://www.eurocommerce.eu/search.aspx?q=Artificial%20Intelligence www.efsa.europa.eu/publications 104 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence • • • • The future EU framework for AI should support the digital development of SMEs The future European framework for AI should rely on a simple, narrow, clear and harmonised definition of AI Careful attention should be paid to avoid overregulation – especially considering recently adopted EU and national legislation, support the use of existing AI technologies and bolster innovation. High-risk AI applications should be defined and overseen in a way that provides legal certainty So, all things considered, EuroCommerce has a favourable view on AI, with few concerns regarding regulations and potential risks. One concern is however presented, since AI could significantly transform the job market, rendering some jobs obsolete. There is also the need to provide public access to a strong supply of digital skills. To sum up, the specific concerns of the stakeholders are summarised in the following table: Table 24 - Stakeholder value map. Stakeholder Category Stakeholder category Sub- Specific Concerns • • Advocacy groups • • • • • NGOs and civil society organisations • Consumer organisations • • • • Practitioners’ associations • • 5.3.2. Transparency, which would allow humans to see whether the AI systems have been thoroughly tested, and provide explanations on why particular decision are made Ethics, trustworthiness and fairness when applying AI in EFSA’s processes, in line with the guidelines of the European Commission and Recommendation 5 “Adopt a Trustworthy AI framework”. Effective accountability systems that would ensure sufficient human oversight and control of automatic decision-making to ensure that human decisionsmakers can be held responsible Right for consumer complaints/legal action in case an AI system or practice affects them, including a right to receive compensation for damages suffered AI practices should not manipulate consumers in a way that could cause them economic harm Right for consumer complaints/legal action in case an AI system or practice affects them, including a right to receive compensation for damages suffered AI should not evaluate individuals on the basis of social behaviours or personal attributes in order to avoid consumer discrimination Transparency, which would allow humans to see whether the AI systems have been thoroughly tested, and provide explanations on why particular decision are made Fairness, which would ensure that unfair bias is avoided, as it might cause multiple negative implications, from the marginalisation of vulnerable groups, to the exacerbation of prejudice and discrimination Effective accountability systems that would ensure sufficient human oversight and control of automatic decision-making to ensure that human decisionsmakers can be held responsible Privacy and data protection of sensible information Transparency, which would allow humans to see whether the AI systems have been thoroughly tested, and provide explanations on why particular decision are made Human centricity, that must ensure that AI applications are centred on the individual, and are always subject to some level of human oversight, especially in decision making processes Effective accountability systems that would ensure sufficient human oversight and control of automatic decision-making to ensure that human decisionsmakers can be held responsible Risks Analysis and Mitigation for Appropriate Communication www.efsa.europa.eu/publications 105 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence On the basis of the main concerns, EFSA should take into consideration the main findings in order to lay down effective communication messages towards its stakeholders, if it is to apply AI in its processes. It is worth noting that these concerns reflect real risks that go hand in hand with the use of AI, which is the reason why, at a policy and regulatory level, the European Union is making a lot of progress to ensure that organisations especially public organisations - adopt AI technology in a trustworthy way. Therefore, failing to address these risks and failing to communicate this effectively to the stakeholders could have negative repercussion towards the agency. Indeed, as a public institution, EFSA must be accountable towards its stakeholders, and must try to avoid being perceived as untransparent in its processes and decisions, which could affect its reputation among its stakeholders. Taking into consideration the main findings of our research, it clear that there are a wide range of needs and concerns from external stakeholders on the use of AI, and therefore external communication of these activities must be well conceived. Indeed, any miscommunication towards these stakeholders – especially to NGOs and civil society organisations - could make them adopt specific actions against EFSA, undermining the Agency’s reputation and perceived transparency, and lowering the accountability of the Agency’s risk assessment activities and decisions. In order to mitigate this risk and to be able to correctly communicate its AI activities externally, we would suggest the following mitigation measures: 1. The correct implementation of the horizontal recommendations, and in particular, the adoption of a Responsible AI framework (no 5) and Definition of Data Governance Framework (no 3). 2. The accurate design of customised and transparent messages that respond to the different specific needs and concerns of the targeted audiences. Table 25 below provides the match between stakeholder sub-categories, their concerns and the topics and concepts that they need to address in order to develop effective communication messages, linking them with the horizontal recommendations. Table 25 - Concepts to integrate into Communication messages. Stakeholders Categories Stakeholders categories Advocacy groups Sub- Stakeholders’ Subcategory Concerns Topics/Concepts to integrate into Communication Messages • • • • Transparency Trustworthy/ Ethical AI Accountability Right to compensation By following Horizontal Recommendation 5 “Adopt a Trustworthy AI framework”, EFSA would ensure the adoption of mechanisms, protocols and guidelines, with a view to: Ensure that humans are aware that they are interacting with an AI system, and are informed of the system’s capabilities and limitations; Adopt AI in a responsible manner while addressing potential ethical risks which might emerged in the maintaining and monitoring of the AI system; Ensure responsibility and accountability for AI systems and their outcomes • Consumer protection • Right to compensation • Consumer discrimination • Transparency • Privacy • Data protection • Need to apply regulations By following Horizontal Recommendation 5 “Adopt a Trustworthy AI framework”, EFSA would ensure the adoption of mechanisms, protocols and guidelines, with a view to: Empower human beings, allowing them to make informed decisions and fostering their fundamental rights; Ensure proper oversight achieved through human-in-the-loop, human-on-the-loop, and human-in-command approaches Avoid unfair bias as it might cause multiple negative implications, from the marginalisation of vulnerable groups, to the exacerbation of prejudice and discrimination Make accessible the system to all, regardless of any disability, and involve relevant stakeholders throughout their entire life circle. NGOs and Civil society Consumer organisations www.efsa.europa.eu/publications 106 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Stakeholders Categories Stakeholders categories Sub- Stakeholders’ Subcategory Concerns Topics/Concepts to integrate into Communication Messages - - Ensure that humans are aware that they are interacting with an AI system, and are informed of the system’s capabilities and limitations; Ensure full respect for privacy and data protection, taking into account the quality and integrity of the data, and ensuring legitimised access to data. In addition, by following Horizontal Recommendation 3 “Implement a Data Governance framework which addresses AI”, EFSA would be looking to: Ensure that data is managed and used correctly and responsibly, while minimising risks, leading to better AI-enabled decision making in general Describe policies, roles, standards, and metrics needed to continuously improve the use of data that ultimately enables an organisation to achieve its goals. Ensure AI compliance with the law Practitioners’ associations • • • • Transparency Human-centricity Accountability Trustworthy/ Ethical AI By following Horizontal Recommendation 5 “Adopt a Trustworthy AI framework”, EFSA would ensure the adoption of mechanisms, protocols and guidelines, with a view to: Empower human beings, allowing them to make informed decisions and fostering their fundamental rights; Adopt AI in a responsible manner while addressing potential ethical risks which might emerged in the maintaining and monitoring of the AI system; Design, develop, and deploy AI systems to learn from and collaborate with humans in a deep, meaningful way Ensure responsibility and accountability for AI systems and their outcomes It is worth noting that EFSA’s messages on trustworthy AI adoption should not be vague, but grounded by tangible actions and results about the way AI processes are evaluated, approved and integrated within EFSA. For this reason, the implementation of Recommendation 5 is particularly important. Otherwise, EFSA could face a major risk of falling into “ethics washing” communication, which is the practice of fabricating or exaggerating an organisation’s interest in an ethical AI system, without being compliant to its principles in a second instance. This could send mixed messages to its stakeholders, and make it lose credibility. 5.4. Engagement Plan To allow a smooth adoption of AI technology in the EFSA evidence management process, a proper engagement plan needs to be set up to target those stakeholders with the highest relevance, and therefore positioned on the right side of the matrix. In this regard, we can distinguish two types of stakeholders: • • [UPPER-RIGHT] Stakeholders that have already collaborated with EFSA in relation to the scope of AI for this project (e.g., Institutions, Academia, Software companies), positioned in the upper right part of the matrix. EFSA should aim to maintain and reinforce the relationship with them, or plan new collaboration opportunities to devise and implement AI projects. [LOWER-RIGHT] Stakeholders (e.g., international organisations, Digital Innovation Hubs), positioned in the lower right of the matrix. They are associated with a high relevance, since collaborating with them could be beneficial for EFSA in view of AI adoption projects. However, they have never collaborated with EFSA on the topic of AI. Therefore, they should be targeted with engagement activities aimed at including them into EFSA's AI ecosystem, in order to foster possible collaboration opportunities. www.efsa.europa.eu/publications 107 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Figure 37 - Maintain & Reinforce Collaboration and Engage in the ecosystem. • It is worth noting that EFSA already has an established Engagement Approach72, in line with its strategy, which can be used with a wide range of stakeholders. The approach follows specific goals, principles and steps, and foresees several mechanisms and tools of engagement and consultation (e.g., discussion groups, information sessions). As EFSA engagement approach provides a general framework to be followed, regardless of the specific domain, we recommend customising this approach, to a certain extent, to the specific scope of this project around the adoption of AI technology for evidence management, to reflect all peculiarities of the AI stakeholders’ ecosystem. In this regard, we recall that section 3.3. describes a list of possible collaboration models to be potentially established with each stakeholder category for each phase of the AI development lifecycle, including modalities to develop and procure innovation, scouting instruments and ideas generation. In this section, we make a step forward by mapping (see Table 26) for each different stakeholder category/subcategory the EFSA main objectives of collaboration for AI adoption (as a result of the Relevance/Impact/Easiness to collaborate assessment), multiple optimal collaboration models to establish such collaboration, the stakeholders’ specific needs (motivation to collaborate) and the tools73 that support EFSA in the engagement model. The analysis proposed in the table below provides the engagement tools that are more likely to be used and more effective in relation to the objective and motivation for the stakeholder to collaborate. In order to choose the most suitable method, EFSA must take into account aspects such as the budget available, the time constraints, the number of resources available and the specificities of this project. For example, Knowledge Fairs are proposed for collaboration models such as Strategic alignment, as they are events designed to share large amounts of information from numerous sources. They are conducted at a common venue and through fishbowls, world cafes or brainstorming in mind mapping exercises, and they are the perfect way to encourage interaction, share experiences, disseminate and promote best practices, and understand peer perspectives. Considering these points, Knowledge Fairs could be a good tool for Strategic Alignment activities, although EFSA should also take into consideration that they would need a significant budget, not only for implementing the event (logistics, travel, support personnel, among others) but also for the preparation phase (experience documentation, promotion, and marketing), as well as for the monitoring phase (knowledge transfer, follow up on agreements or statements, mobilisation of communities of practice, etc) 74. This argument can be applied also to the remaining engagement tools provided in Table 26, meaning that EFSA must think about which will be the most effective and appropriate tools for each specific occasion. 72https://www.efsa.europa.eu/sites/default/files/EFSA%20Stakeholder%20engagement%20approach_FINAL.pdf 73 To map the tools, the EFSA stakeholder engagement toolkit was consulted https://www.efsa.europa.eu/sites/default/files/documents/engagement-toolkit.pdf 74 https://www.shareweb.ch/site/Learning-and-Networking/sdc_km_tools/Documents/Knowledge%20Fair_Undp.pdf www.efsa.europa.eu/publications 108 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence www.efsa.europa.eu/publications 109 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Table 26 - Engagement plan. Stakeholder Category Stakeholder Sub-Category EFSA Main Objective Collaboration/Coop eration Models Needs and Added value for EFSA to Collaborate Motivation collaborate of stakeholders to Strategy: European Commission Maintain Reinforce Collaboration & o Align on existing and planned AI strategies, principles, and policies of different actors Ecosystem: Strategic Alignment o Sharing of plans, knowledge, lessons learned and best practices o Learn policies and standards for ethical use of AI from other organisations o Line up AI use cases and opportunities from different actors o Set grounds for future cooperation along the following phases of the AI lifecycle Strategic Alignment Ecosystem: EU Agencies Maintain Reinforce Collaboration & o Learn policies and standards for ethical use of AI from other organisations o Sharing of plans, knowledge, lessons learned and best practices o Opportunity for future collaboration for AI adoption o Scouting for new technologies options o Alignment and/or collection of feedback on the Technology roadmap development Design: o Participate in a joint action with shared budget, governance and risks for the acquisition or co-development of AI solutions o Implementation o Participate in a joint action with shared budget, governance and risks for the deployment and integration of AI solutions • Public events (Workshops, Webinars, Technical meetings) • Discussion Groups and Forums • Stakeholder Forum • Focus Group • Online Platforms 75 • Vision Factory • PESTEL and SWOT analysis • Open space technology76 Info-sessions, • Envisioning the future 77 • Knowledge Fair • Conferences Strategy: o Align on existing and planned AI strategies, principles, and policies of different actors Institutions Collaboration/Cooperation Tools Cooperation programme o Creation of synergies, sharing best practices, lessons learned and resources o Shared risk and costs • Public events (Workshops, Webinars, Technical meetings) • Discussion Groups and Forums • Stakeholder Forum • Focus Group • Online Platforms • Vision Factory • PESTEL and SWOT analysis • Open space technology • Envisioning the future • Knowledge Fair • Conferences Info-sessions, • Public events (Workshops, Webinars, Technical meetings) • Focus Group, Interviews • Brainstorming • Reverse Brainstorming • Online platform • Consensus building processes Info-sessions, 75 It is a combination of World Cafés for the future development of complex topics, which can be used, for instance, to create engagement and consensus on a joint vision. Open Space Technologies is a method to organise participation events, basically of large and medium scale. It can be used to support informal learning, brainstorming, networking, deal making and collaboration within groups that have identified common goals; address highly complex central themes that no single person or small group can understand completely; design action plans. 77 It is a scenario-building method that invites collective reflection about plausible futures. It works by imagining a time in the future (three to six years ahead), and assumes that the organisation, section or field presence has achieved important goals. 76 www.efsa.europa.eu/publications 110 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence Stakeholder Sub-Category EFSA Main Objective Needs and Added value for EFSA to Collaborate Collaboration/Coop eration Models Motivation collaborate of stakeholders to Collaboration/Cooperation Tools • Ad hoc working group • After action review (AAR)78 Design: o Access to European funds to undertake R&D projects with other organisations to design innovative solutions Funding Projects (e.g., Horizon Europe) Design: o Creation of synergies, sharing best practices, lessons learned and resources o Shared risk and costs o Visibility Contest for prize o Contribution to the research and development of innovative products and services o Line up AI use cases and opportunities from different actors the o Set grounds for future cooperation along the following phases of the AI lifecycle Strategic Alignment Ecosystem: o Learn policies and standards for ethical use of AI from other organisations o Sharing of plans, knowledge, lessons learned and best practices o Opportunity for future collaboration for AI adoption o Scouting for new technologies options o Alignment and/or collection of feedback on the Technology roadmap development Design: National Competent Authorities (incl. Art. 36 org) Maintain Reinforce Collaboration & o Delegate the design of innovative solutions through the release of a grant, through effective and efficient partnership with competent organisations that are active in the field of EFSA, and are designated by the respective Members States • Online platform • Website o Reward and/or profit Strategy: o Align on existing and planned AI strategies, principles, and policies of different actors Engage in ecosystem Info-sessions, o Reputation o Outsource to different third parties the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems International Organisations • Public events (Workshops, Webinars, Technical meetings) • Focus Group, Interviews • Brainstorming • Reverse Brainstorming • Online platform • Consensus building processes • Ad hoc working group • After action review (AAR) o Creation of future opportunities for collaboration Grant Implementation: o Strengthen relationship with EFSA o Easiness to collaborate (i.e., reduced administrative burden) • Public events (Workshops, Webinars, Technical meetings) • Discussion Groups and Forums • Stakeholder Forum • Focus Group • Online Platforms • Vision Factory • PESTEL and SWOT analysis • Open space technology • Envisioning the future • Knowledge Fair • Conferences Info-sessions, • Technical Meetings, Workshops • Design Thinking Project79 • Brainstorming • Reverse Brainstorming • Concept mapping80 78 It is a simple process to review a project, an activity, an event or a task. In an AAR, the individuals involved discuss what happened, why it happened, what went well, what needs improvement, and what lessons can be learned from the experience, with a view to doing as well or better next time. 79 It is a design methodology that provides a solution-based approach to solving problems. It is extremely useful in tackling complex problems that are ill-defined or unknown, by understanding the human needs involved, and by re-framing the problem in human-centric ways. This is pursued by stimulating the creation of innovative ideas in brainstorming sessions, and by adopting a hands-on approach in prototyping and testing through the five stage of Design Thinking. 80 Concept mapping is a structured process, focused on a topic or construct of interest, involving input from one or more participants, that produces an interpretable pictorial view (concept map) of their ideas and concepts, and how these are interrelated. www.efsa.europa.eu/publications 111 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Stakeholder Category Stakeholder Sub-Category EFSA Main Objective Needs and Added value for EFSA to Collaborate Collaboration/Coop eration Models Motivation collaborate of stakeholders to o Delegate the deployment and transition of innovative solutions through the release of a grant, through effective and efficient partnership with competent organisations that are active in the field of EFSA, and are designated by the respective Members States Design: o Access to European funds to undertake R&D projects with other organisations to design innovative solutions Collaboration/Cooperation Tools User committee81 Online Platforms After action review (AAR) Participatory design Interviews, Survey, Focus group, webinars (for implementation • Group interview with a co-design session82 • • • • • Funding Projects (e.g., Horizon Europe) Design: o Creation of synergies, sharing best practices, lessons learned and resources o Shared risk and costs o Visibility • Public events (Workshops, Webinars, Technical meetings) • Focus Group, Interviews • Brainstorming • Reverse Brainstorming • Online platform • Consensus building processes • Ad hoc working group • After action review (AAR) Info-sessions, o Reputation o Outsource to different third parties the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems Contest for prize o Contribution to the research and development of innovative products and services • Online platform • Website o Reward and/or profit Data: o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., definition of suitable data requirements) to gain efficiencies and flexibility Implementation o Outsource to a third-party the deployment and transition of an off-the-shelf AI solution that already meets the requirements and needs, in a cost and time effective way with relatively low risks Operating o Funds to pursue their activities Procurement o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., definition of services levels, incident, and problem management procedures) to gain efficiencies and flexibility o Establishment of a potential long-term relationship o Reputation • • • • Technical Meetings, Workshops After action review (AAR) Online Platforms Interviews, Survey, Focus group, webinars (for implementation) Monitoring o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., performance monitoring of the solution) to gain efficiencies and flexibility Industry 81 82 Software companies Maintain Reinforce Collaboration & Procurement Data: o Profit • Technical Meetings, Workshops • After action review (AAR) • Online Platforms This method involves users and other stakeholders in the formal monitoring and steering of the research and innovation process. The group interview with a co-design session will provide feedback about the research scenarios presented. www.efsa.europa.eu/publications 112 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Stakeholder Category Stakeholder Sub-Category EFSA Main Objective Needs and Added value for EFSA to Collaborate Collaboration/Coop eration Models o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., definition of suitable data requirements) to gain efficiencies and flexibility Motivation collaborate of stakeholders to o Establishment of a potential long-term relationship Collaboration/Cooperation Tools • Interviews, Survey, Focus group, webinars (for implementation) o Reputation Implementation o Outsource to a third-party the deployment and transition of an off-the-shelf AI solution that already meets the requirements and needs, in a cost and time effective way with relatively low risks Operating o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., definition of services levels, incident, and problem management procedures) to gain efficiencies and flexibility Monitoring o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., performance monitoring of the solution) to gain efficiencies and flexibility Design: o Outsource to a third-party the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems o Request for a demonstration of feasibility and quality of the solution, through a Proof of Concept o Profit Innovation Procurement o Establishment of a potential long-term relationship o Reputation Implementation: Outsource to a third-party the deployment and transition of an innovative AI solution in the market, aimed at improving public services effectiveness and efficiency, and solving new problems Design: o Access to European funds to undertake R&D projects with other organisations to design innovative solutions Funding Projects (e.g., Horizon Europe) Design: Contest for prize www.efsa.europa.eu/publications the o Visibility • Public events (Workshops, Webinars, Technical meetings) • Focus Group, Interviews • Brainstorming • Reverse Brainstorming • Online platform • Consensus building processes • Ad hoc working group • After action review (AAR) Info-sessions, o Outsource to a third-party for non-competitive use cases (for which a start-up tends to perform better than a stand-alone solution for one specific application in one company) the design 113 o Contribution to the research and development of innovative products and services • Online platform • Website o Reward and/or profit Design: Engage in ecosystem o Shared risk and costs Technical Meetings, Workshops Design Thinking Project Brainstorming Reverse Brainstorming Concept mapping User committee Online Platforms After action review (AAR) Participatory design Interviews, Survey, Focus group, webinars (for implementation o Reputation o Outsource to different third parties the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems AI startups o Creation of synergies, sharing best practices, lessons learned and resources • • • • • • • • • • Innovation Procurement o Profit o Establishment of a potential long-term relationship • • • • Technical Meetings, Workshops Design Thinking Project Brainstorming Reverse Brainstorming EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Stakeholder Category Stakeholder Sub-Category EFSA Main Objective Needs and Added value for EFSA to Collaborate Collaboration/Coop eration Models and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems Motivation collaborate of stakeholders to o Reputation o Growth and stability o Request for a demonstration of feasibility and quality of the solution, through a Proof of Concept Implementation: Collaboration/Cooperation Tools • • • • • • • Concept mapping User committee Online Platforms After action review (AAR) Participatory design Innovation Jam Interviews, Survey, Focus group, webinars (for implementation o Outsource to a third-party the agile deployment and transition of an innovative AI solution in the market, aimed at improving public services effectiveness and efficiency, and solving new problems o Reputation Design: o Outsource to different third parties the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems Contest for prize o Contribution to the research and development of innovative products and services • Online platform • Website o Reward and/or profit o Growth and stability Design: o Outsource to a third-party for non-competitive use cases (for which a start-up tends to perform better than a stand-alone solution for one specific application in one company) the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems o Request for a demonstration of feasibility and quality of the solution, through a Proof of Concept University offs Spin- Engage in ecosystem the o Profit Innovation Procurement o Establishment of a potential long-term relationship o Reputation o Growth and stability Implementation: • • • • • • • • • • Outsource to a third-party the agile deployment and transition of an innovative AI solution in the market, aimed at improving public services effectiveness and efficiency, and solving new problems Technical Meetings, Workshops Design Thinking Project Brainstorming Reverse Brainstorming Concept mapping User committee Online Platforms After action review (AAR) Participatory design Interviews, Survey, Focus group, webinars (for implementation o Reputation Design: Outsource to different third parties the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems Contest for prize o Contribution to the research and development of innovative products and services • Online platform • Website o Reward and/or profit o Growth and stability Data: Academia Universities and research centres (incl. Art. 36 org) Maintain Reinforce Collaboration & o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., definition of suitable data requirements) to gain efficiencies and flexibility o Funds to pursue their activities Procurement Implementation o Outsource to a third-party the deployment and transition of an off-the-shelf AI solution that already meets the requirements and needs, in a cost and time effective way with relatively low risks o Establishment of a potential long-term relationship o Reputation • • • • Technical Meetings, Workshops After action review (AAR) Online Platforms Interviews, Survey, Focus group, webinars (for implementation) Operating www.efsa.europa.eu/publications 114 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Stakeholder Category Stakeholder Sub-Category EFSA Main Objective Needs and Added value for EFSA to Collaborate Collaboration/Coop eration Models Motivation collaborate of stakeholders to Collaboration/Cooperation Tools o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., definition of services levels, incident, and problem management procedures) to gain efficiencies and flexibility Monitoring o (Partially) outsource to a third-party non-core activities or activities for which skills are not fully developed internally (e.g., performance monitoring of the solution) to gain efficiencies and flexibility Design: o Delegate the design of innovative solutions through the release of a grant, through effective and efficient partnership with competent organisations that are active in the field of EFSA, and are designated by the respective Members States Implementation: o Creation of future opportunities for collaboration Grant Design: o Access to European funds to undertake R&D projects with other organisations to design innovative solutions o Strengthen relationship with EFSA o Easiness to collaborate (i.e., reduced administrative burden) o Delegate the deployment and transition of innovative solutions through the release of a grant, through effective and efficient partnership with competent organisations that are active in the field of EFSA, and are designated by the respective Members States Funding Projects (e.g., Horizon Europe) o Creation of synergies, sharing best practices, lessons learned and resources o Shared risk and costs o Visibility Design: o Outsource to a third-party the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems o Request for a demonstration of feasibility and quality of the solution, through a Proof of Concept o Funds to pursue their activities Innovation Procurement o Establishment of a potential long-term relationship o Reputation Implementation: o Outsource to a third-party the deployment and transition of an innovative AI solution in the market, aimed at improving public services effectiveness and efficiency, and solving new problems Design: Technical Meetings, Workshops Design Thinking Project Brainstorming Reverse Brainstorming Concept mapping User committee Online Platforms After action review (AAR) Participatory design Interviews, Survey, Focus group, webinars (for implementation • Group interview with a co-design session • Public events (Workshops, Webinars, Technical meetings) • Focus Group, Interviews • Brainstorming • Reverse Brainstorming • Online platform • Consensus building processes • Ad hoc working group • After action review (AAR) • • • • • • • • • • Info-sessions, Technical Meetings, Workshops Design Thinking Project Brainstorming Reverse Brainstorming Concept mapping User committee Online Platforms After action review (AAR) Participatory design Interviews, Survey, Focus group, webinars (for implementation o Reputation o Outsource to different third parties the design and testing of an innovative AI solution in the market, for which the organisation has not the resources and/or competences, aimed at improving public services effectiveness and efficiency, and solving new problems www.efsa.europa.eu/publications • • • • • • • • • • 115 Contest for prize o Contribution to the research and development of innovative products and services • Online platform • Website o Reward and/or profit EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Stakeholder Category Stakeholder Sub-Category EFSA Main Objective Needs and Added value for EFSA to Collaborate Collaboration/Coop eration Models the stakeholders to o Alignment and/or collection of feedback on the Technology roadmap development o Widen the engagement outreach of EFSA towards various stakeholders, whenever a procurement, contest for prize or grant etc. is launched in the context of AI adoption, in order to create synergies Info-sessions, Contact • Public events (Workshops, Webinars, Technical meetings) • Online Platforms • Open space technology • Conferences • Brainstorming • Reverse Brainstorming • Concept mapping • E-mails • Social Media/website Info-sessions, o Assist technology transfer by providing networking and advocacy, and finding a qualified commercial partnership to successfully and diligently develop a technology discovered and researched at a university and research institutions Contact • Public events (Workshops, Webinars, Technical meetings) • Online Platforms • Open space technology • Conferences • Brainstorming • Reverse Brainstorming • Concept mapping • E-mails • Social Media/website Info-sessions, o Point of contact by providing strong linkages with different actors inside and outside of its region, promoting the implementation of multiple digital opportunities in the industry and the public sector, sustainability of their processes and products o Support to public sector organisations in the use of digital technology Contact Consultation and o Point of contact by providing strong linkages with different actors inside and outside of its region, promoting the implementation of multiple digital opportunities in the industry and the public sector, sustainability of their processes and products Design and Implementation Development Levers Technology transfer office Engage in ecosystem the o Widen the engagement outreach of EFSA towards various stakeholders, whenever a procurement, contest for prize or grant etc. is launched in the context of AI adoption. o Establish a partnership with Academia that has discovered and researched ground-breaking technologies o Access to knowledge, bringing skills and experience to the surface, which are necessary tools for the implementation of EFSA aspirations Incubators and accelerators Engage in ecosystem www.efsa.europa.eu/publications the Ecosystem o Widen the engagement outreach of EFSA towards various stakeholders, whenever a procurement, contest for prize or grant etc. is launched in the context of AI adoption. o Access to knowledge, bringing skills and experience to the surface, which are necessary tools for the implementation of EFSA aspirations 116 Collaboration/Cooperation Tools • Public events (Workshops, Webinars, Technical meetings) • Online Platforms • Open space technology • Conferences • Brainstorming • Reverse Brainstorming • Concept mapping • E-mails • Social Media/website o Scouting for new technologies options Engage in ecosystem of o Provision of access to technical expertise and experimentation needed for a successful digital transformation. Ecosystem: Digital Innovation Hubs Motivation collaborate EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Stakeholder Category In addition, the following table shows other cross-cutting collaboration/cooperation models with a wide outreach and related tools, which can be employed for any stakeholder category/sub-category: Table 27 - Collaboration models and related tools Collaboration/Cooperation Models Public consultation Internet crowdsourcing Hackathon 5.4.1. Needs and Added value for EFSA to Collaborate Stakeholders’ Motivation to collaborate Strategy: • Understand the opinion of the ecosystem’s actors on the adoption of AI strategies and principles within organisations to guide decision-making Strategy: • Access to a large audience with different expertise to collect new ideas and opportunities on the AI topic Ecosystem: • Scouting for new technologies options Design: • Generate ideas for new AI solutions and services, stimulating the ecosystem of actors • • • • • • Contribution to the technology transformation of public sector Opportunity for future collaboration for AI adoption Contribution to the technology transformation of public sector Monetary or non-monetary awards Contribution to the technology transformation of the public sector Experience gain Collaboration/Cooperation Tools • IT tools Connect.EFSA) • Online Platforms Innocentive) • • Open Platforms Design Thinking Project (via (e.g. Communication on AI adoption Although our approach suggests that EFSA should set up an engagement plan for the stakeholders placed in the right part of the matrix (to engage new actors in the ecosystem or develop/strengthen collaboration with existing actors), this does not exclude the parallel need to foresee some communication activities (through the methods and channels EFSA has available) towards these actors, regarding EFSA’s strategy, objectives, plans and activities on the adoption of AI for evidence management. This is important to spread awareness among stakeholders, and to sow the ground for future collaboration opportunities. Moreover, to promote certain modes of collaboration it would be necessary to adopt more specific communication activities, as in the case of the calls for the purchase of AI solutions, or the contests for a prize, addressing specific stakeholders’ categories interested directly or indirectly (i.e., informed through development levers, e.g. DIHs). 5.5. Final Considerations on Communication and Engagement – Timeline and dependencies To conclude, we outline some final considerations on the correct time and existing dependencies of the communication and engagement plans with other recommendations for the implementation of the communication plan and engagement plan. • • Communication Plan: it should be launched once Horizontal Recommendation 5 on the Trustworthy AI framework is correctly implemented (or at least planned and launched), to avoid the risk of ethics washing. For this reason, the first step of the Recommendation’s activities (namely, the development of the Code of Conduct) should start quite soon, as shown in Table 18. Engagement Plan: this can start immediately based on the needs, prior to and during the adoption of AI technology, in order to reinforce, widen and cooperate with EFSA's AI ecosystem. Therefore it is not subject to any particular dependency with other recommendations or constraints. www.efsa.europa.eu/publications 117 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence In addition, it is worth considering that these plans can be implemented both at a macro and a micro level. In the first case, the plans relate to an entire programme, which contains multiple AI projects (horizontal and vertical recommendations), and therefore said plans envisage more generic and strategic communication and engagement activities, with a long-term strategic vision on the adoption of AI for EFSA's evidence management. In the second case, the activities of the communication and engagement plans are tailored to the specific needs of the AI projects (e.g. organisation of a workshop with another EU agency for aligning on the development of an AI trustworthy framework). 6. Conclusions The output of this roadmap will serve as a decision-making tool to identify and prioritise the AI projects to be carried out from 2022 to 2027, and to optimise the performance of evidence management through AI in order to achieve the objectives presented in EFSA's theme paper. By implementing the projects devised in this roadmap, EFSA will have developed the fundamental capabilities that will enable the integration of AI within the organisation, building not only the technical instruments to manage, control, and use AI responsibly, but also in terms of skills and cultural change. In addition, EFSA will have already integrated some AI solutions in several use cases, resulting in a more efficient, performative and automated evidence management process 83. These projects will be particularly relevant for building the Agency’s capabilities to integrate AI tools inside the evidence management process, but also in other processes, thus forging synergies with the needs of other roadmaps (e.g., the EU Partnership for next generation, systems-based environmental risk assessment - PERA - and the New Approach Methodologies - NAMs), and of the organisation’s readiness towards AI adoption as a whole. Table 28 - Focus on synergies with other roadmaps. New Approach Methodologies (NAMs roadmap) A major element of the application of New Approach Methodologies (NAMs) is the need for integrated approaches for testing and assessment (IATAs), defined approaches for data interpretation, and performance-based evaluation of test methods. An issue identified by the NAMs project team is the need for improved data-integration tools to facilitate the gathering of multiple disparate data sources and the transformation of these data sources into standardised format for processing. While no vertical recommendations were identified as relevant to the NAMs roadmap, Horizontal recommendations around the Development and adoption of a Data Governance, Data Architecture for Big Data and AI are both recognised as foundational elements for building the appropriate infrastructure and tools to achieve improved data extraction, integration and standardisation to support the use of NAMs in chemical risk assessment. In addition, once AI techniques are applied in NAMs for decision support in risk assessment, the adoption of a Responsible AI framework will also be of paramount importance. EU Partnership for next generation, systems-based Environmental Risk Assessment (PERA roadmap) 83 As already explained in the previous chapters, with regard to the 10 prioritised use cases, the optimal vender scenario for developing or acquiring tailored AI-solutions (vertical recommendations) has been already selected thanks to the market scouting performed during the data collection activities of the project. However, the vendor selection approach developed in section 3.3.2.1 can be applied to all AI use cases, and therefore EFSA might use it to identify the best way to adopt AI solutions also in other – not prioritised – use cases. www.efsa.europa.eu/publications 118 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence One of the major issues of the PERA project is the resource-intensiveness when retrieving and analysing relevant information from databases and websites. Use cases included in our vertical recommendations that can introduce time-efficiencies in this field are the SLR Literature deduplication, SLR Relevant Abstract Screening, SLR Keyword Identification and Automatic Text Summarisation. An additional pain point of PERA is the variability between the outcomes of peer-reviews conducted by different risk assessors. AI can help in this case through the application of topic modelling approaches or Automatic Text Summarisation, which can help simplify the procedure by helping PERA to produce informative summaries per each peer-review. These summaries will contain key elements of the review, which can be used in an AI tool (such as Comparable Appraisal Detection) to identify similarities and discrepancies. Lastly, PERA team suggested that an AI tool that will be able to evaluate how well an interview has been transcribed into text format could really benefit in terms of time and workload reduction. PERA has its own database of texts that have already been transcribed, and therefore this can be used as “training data” for the AI tool, and to further increase its accuracy. Modifying slightly the use case of Data Collection Terminology Assessment would help address the challenge, since the logic of assessing the text format and terms is the same. Thanks to the results that will be achieved by implementing these projects, it will be possible and appropriate after 2027 to rethink the evidence management process as a whole, redesigning it in a more integrated way to leverage the full potential of AI technology (with a technology-driven approach), for example, eliminating steps and sub-processes that are no longer necessary, or merging steps that were previously separated and could be carried out simultaneously. To fill the gap between “adopting and integrating AI” and “embedding AI” in the organisation, thus making it not only an instrument of automation but also a means of holistic innovation, can be considered as the objective to achieve in the long run. In the meantime, EFSA should monitor the key emerging technological trends that will shape the future public sector innovation scenario, which might further boost a smooth adoption of AI within the organisation. In addition, EFSA should also focus attention on its stakeholder ecosystem, keeping the most relevant actors engaged and aligned on the future outcomes of the roadmap and on future activities, with the aim of creating a flourishing AI community around EFSA, and building solid partnerships and collaborations. Concerning the engagement of the AI stakeholder ecosystem, it is worth considering that the stakeholders that we called “innovation development levers” (i.e., accelerators, digital innovation hubs), as presented in section 3.3.2, will play a more impactful and relevant role in the years to come, due to an increase in EU and national funding for their development. In fact, building on and complementing various national initiatives within the digital industry, the Commission is acting to trigger further investments in the digitisation of industry, and thereby support the creation of better framework conditions for the digital industrial revolution. For instance, the EC is further developing the network of (European) Digital Innovation Hubs through the Digital Europe Programme, which will increase the innovation potential of European SMEs, also with a focus on specific technologies such as AI [1]. At national level, investments are being made in several EU Member States as well to develop accelerators and incubators. These kinds of organisations can provide EFSA with strong linkages with innovative companies inside and outside of its region, promoting public services innovation and providing an exchange environment where calls for ideas and innovation procurement can be launched, drawing on the knowledge, approaches and skills of startups, thereby contributing to EU digitalisation by driving demand through open innovation. Therefore, although in the short run these types of stakeholders will [1] https://digital-strategy.ec.europa.eu/en/activities/edihs www.efsa.europa.eu/publications 119 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence probably have a minor role, in the middle term they will have built the capacity to be an important part of EFSA's AI stakeholders’ ecosystem. Another emerging trend to consider in the future - after 2027 - is the European AI On Demand Platform and Ecosystem84, developed by the AI4EU consortium (established in January 2019) with the support of the European Commission under the H2020 programme. The platform is a one-stop-shop for anyone looking for AI knowledge, technology, tools, services and experts, and aims to increase innovation and technology transfer, accelerate the growth of start-ups and SMEs, and fulfil the needs of the European AI community. It will enable access to a catalogue of AI-based resources and a marketplace providing central, easy and simple access to trustworthy AI tools developed in Europe. This will be another asset that EFSA can leverage, supporting its own process towards the adoption of AI and the implementation of the Adopt AI programme. In the next 5 years leading up to 2027, EFSA can follow a structured guideline/roadmap for actions, with key associated tools, to adopt AI for evidence management. Using the adoption of AI for evidence management as a reference and best practice for other organisational processes, EFSA may build the fundamental capabilities for an appropriate, organisation-wide AI implementation. 84 https://www.ai4europe.eu/ www.efsa.europa.eu/publications 120 EFSA Supporting publication 2022: EN-7339 The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors. 23978325, 2022, 5, Downloaded from https://efsa.onlinelibrary.wiley.com/doi/10.2903/sp.efsa.2022.EN-7339, Wiley Online Library on [30/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Roadmap for actions on artificial intelligence