Using generative AI to turbocharge digital marketing Received (in revised form): 26th September, 2023 Ian Thomas Founder/Chief Data Officer, Yew Tree Data Consulting, UK Ian Thomas cofounded one of the industry’s first Web analytics firms and has held senior data leadership roles for Microsoft, Publicis Groupe and Omnicom. He has managed some of the world’s largest and most complex datasets and built effective cross-functional teams to bring data to life. He now spends time working with CDOs and other senior data leaders to help them transform the value they derive from their investments in data, analytics and AI. E-mail: ianth@outlook.com; Website: https://www.yewtreedata.com Abstract AI has been used in marketing for some time to enable better targeting and optimisation of messages, but some of the benefits of these approaches have been limited by the ability to create personalised content at scale and the ability to measure the effectiveness of this content in a structured way. Generative AI offers a way to address these challenges and to facilitate integrations that make it easier to execute and measure marketing in a fragmented ecosystem. However, the technology presents a series of technical, privacy and copyright challenges that organisations will need to overcome in order to use it effectively. KEYWORDS: generative AI, OpenAI, digital marketing, GDPR, LLMs, privacy, personalisation GENERATIVE AI IN DIGITAL MARKETING The use of artificial intelligence (AI) is well established in digital marketing. As the availability of rich audience data grew in the 2010s, advertisers and publishers used this data to slice audiences into finer and finer segments and delivered targeted campaigns and creatives to these segments. Machine learning (a key AI technique) was used both to create the segments and decide which messages to deliver to whom, using a combination of classification, prediction and recommendation algorithms (Table 1). As audience data has become harder to acquire and work with, creative/message optimisation has become more important. Treating content as its own experimentation 270 space and optimising variations in tone, emphasis and message has the potential to drive significant performance improvements for marketing without the need to exploit personal data. This is where generative AI techniques have a valuable role to play. How generative AI differs from other forms of AI Generative AI (GAI) builds on a set of established techniques within the AI discipline, most notably deep learning, but has a unique emphasis: rather than aiming to make automated decisions or predictions, GAI focuses on producing original content. GAI models are trained on very large amounts of publicly available data. They build a complex statistical model of the Applied Marketing Analytics Vol. 9, 3 270–280 © Henry Stewart Publications 2054-7544 (2023) (2023) Using generative AI to turbocharge digital marketing Table 1: AI system types and their applications in digital marketing AI system type Example applications in digital marketing Classifiers (‘what kind of thing is this?’) • Audience segmentation • Creative asset classification • Creative whitespace identification • Lead qualification • Lookalikes Predictors (‘what is going to happen?’) • Target segment selection • Campaign ROI prediction • Media mix modelling • Creative/keyword selection • Bid selection Recommenders (‘this is what you should do next’) • Product recommendation • Cross-sell/up-sell Generators (‘Hello, Dave . . .’) • Automated creative/copy generation • Creative classification and feature extraction • Chatbots • Content localisation • MarTech integration automation characteristics of the input data, which enables them to generate new content that resembles that on which they were trained. Although there has appeared to be a sudden set of breakthroughs in GAI in the last year, leading to its sudden explosion as a topic of interest, the advances in this field are based on a linked set of innovations stretching back over the last decade. For example, transformer architectures, a key element of many generative models, were first developed and explained in a paper from Google in 2017,1 while generative adversarial networks, which have been essential to the development of image-based GAI, were first introduced in 2014.2 For several years, as these techniques were refined by organisations like Google and OpenAI, progress was impressive but not transformative. This changed in 2022 when OpenAI released its GPT-3 large language model (LLM), a transformer-based model that was trained on 175 billion parameters,3 an almost hundredfold increase on the previous version, which resulted in a quantum leap in performance across a number of scenarios such as translation, summarisation and question answering. Since then, GPT-4 has been released; the number of parameters used to train it is thought to be around a trillion,4 representing another hundredfold increase from GPT-3. Training these models with such a large amount of data has been very expensive: it is believed that GPT-4 cost US$100m to train. This leads to a key difference between GAI models and other kinds of machine learning models, which is that most organisations cannot afford to train their own GAI model from scratch. Instead, the usage pattern is for organisations to make use of so-called ‘foundational models’ and to tune (or guide) these models to be able to apply them to specific situations. GENERATIVE AI MARKETING USE CASES A lot of discussion on GAI in a marketing context has focused on its ability to accelerate and automate the creative/content generation process, but there are a number of valuable use cases beyond this. Creating scaled content variation The rich availability of audience and contextual data has enabled the creation of © Henry Stewart Publications 2054-7544 (2023) Vol. 9, 3 270–280 Applied Marketing Analytics 271 Thomas sophisticated dynamic creative optimisation (DCO) solutions which combine creative elements to create almost limitless variation in creative. Despite these tools’ power, however, they still mostly rely on the raw components and their variants being created manually, and they work better for advertising use cases than, for example, e-mail. GAI can provide additional capability here by creating content variations (especially textual ones) which convey the same information but vary in their tone, length and emphasis of key points. One of the key limitations here for GAI solutions is that GAI models struggle to maintain fidelity to reference assets such as brand imagery, and they cannot be relied on not to violate brand guidelines without supervision. For example, Stability.ai’s DreamStudio tool can take an image and generate variants of it using the Stable Diffusion algorithm. Starting with an image of the Nike logo, and using the prompt ‘the nike logo in beautiful colors surrounded by exotic flowers’ generated the outputs shown in Figure 1. As can be seen from the generated images, the model has modified the Nike logo — even the third example is slightly modified and would therefore not pass muster from a brand perspective. For this reason, GAI functionality is likely to be incorporated into more structured DCO and content generation toolsets, such as Adobe’s Firefly technology, which is used to add generative capabilities to tools like Photoshop and Illustrator. Given the (comparatively) high cost of generating creative variants with GAI (versus just using combinatorial techniques), it is likely that a blended approach will be most valuable, where GAI is used for message ideation and a more workaday combinatorial approach used for variant generation. This approach is exemplified by Adcreative.­ai, an online creative generation tool primarily aimed at small to medium businesses who need to create new campaign messaging, and variants on that messaging, quickly. Larger advertisers (and their agencies) may use separate toolsets for these two parts of the creative generation process (Table 2). Unsurprisingly, the major players in the advertising industry have announced that they will be providing GAI-enabled creative tooling for their platforms. Meta has announced5 that it will be debuting generative capabilities in its ad platform by the end of 2023, and Google has also trailed similar functionality in its platform.6 Creative classification and feature extraction As automated creative generation has taken off, one of the challenges it has created is in measurement — if many variants of an advert are used in a campaign, it is essential to know which ones the audience is responding do. Historically, creative agencies Figure 1: Nike logo variations generated by Stability.­ai 272 Applied Marketing Analytics Vol. 9, 3 270–280 © Henry Stewart Publications 2054-7544 (2023) Using generative AI to turbocharge digital marketing Table 2: GAI/DCO strengths and weaknesses Generative AI DCO Good for • Accelerating creative origination using prompts/reference material • Generating natural language variants that are not template-based • Identifying ‘rich context’ — eg parsing web page content and using this to customise advert copy • Creating structured variants on core creative • Generating variants for segmentation/ targeting (either at user level or contextual level) • Brand safety (working with and staying faithful to existing brand assets) Less good at • Working with/recombining existing creative assets • Generating structured data for analytics • Coming up with ‘new’ ideas • Incorporating rich context and media agencies have not worked together well to create this layer of metadata that is necessary to categorise creative variants, especially along multiple dimensions, which makes it harder to understand which creatives worked well, which did not and why that might be. One of the key features of GAI models is their ability to create a simplified representation of complex textual, image or video information. This technique can be used to take a set of creative assets and create structured metadata about them that can be used to classify them, aiding in subsequent performance analysis. Even better, such attribute extraction can be done on a post-hoc basis using categories that were not thought of when the creative was created. Localisation The current generation of large language models are very good at machine translation across the more popular languages, with an increasingly impressive grasp of idiom, which makes them practical for localisation of advertising creative at scale, especially in more utilitarian contexts (where the language is relatively simple). Additionally, image GAI models can also create localised variants of visuals for adverts to align with local culture and preferences. Combined with the ability to easily create multiple variants of creative, this means that brands can scale their campaigns internationally much more easily and select the creative that works best in each market or region. Recommenders and personalisation The core techniques used in GAI models makes them quite well suited to recommendation tasks, especially over relatively bounded sets of options. Again, the technique of creating embeddings, which is foundational to GAI models, is of great use in recommender scenarios since it enables identification of items (such as products) that are similar to each other across a range of dimensions (which can include the characteristics of people who buy those products, or other products that are also bought alongside). Although it is possible to build a recommender system on top of a commercial large language model (such as GPT or PaLM) by feeding information about the user and the available options/ products into the prompt, this is likely to be a very inefficient way to address this need at anything below low volumes, and it runs into significant privacy issues. A more compelling use case for LLMs here is to combine the recommendation with content generation — for example, weaving recommendations into a personalised marketing e-mail alongside other content. © Henry Stewart Publications 2054-7544 (2023) Vol. 9, 3 270–280 Applied Marketing Analytics 273 Thomas MarTech platform integration A major use case for large language models is the generation of code. Commercial platforms like GitHub are already integrating ‘copilot’ functionality that can generate code based on textual prompts and to complete existing code. This dramatically reduces the friction of creating new code, especially for ‘utility’ scenarios such as the creation of integration code that uses a published API. The fragmented nature of the MarTech/AdTech ecosystem means that integration of disparate platforms is a major challenge for organisations, so the automation of this process by GAI tools is likely to have a major impact. Low-code integration vendors like Zapier are already taking advantage of GAI technology in their own platforms.7 Natural language analytics A special case of text-to-code generation is the creation of SQL (or Python) code that can be used to retrieve and return data to answer a specific question. This has the potential to transform how individuals interact with data and may be the tipping point for driving much greater data literacy and usage within the organisation. However, these tools still make both obvious and more subtle mistakes in the code they generate,8 which creates significant risk that non-data specialist users will accidentally create misleading or inaccurate numbers with these Table 3: 274 kinds of tools. Specialist analytics personnel will probably still be needed for some time to come to prevent these mistakes from occurring. GENERATIVE AI IMPLEMENTATION MODES Organisations that wish to utilise GAI models have three major options available at present, which provide differing sets of benefits: • direct usage of a public model (eg ChatGPT) via interface or API; • usage of a dedicated instance of a foundational model, hosted in public cloud, with or without model tuning; • usage of an on-prem or private cloud instantiation of a foundational model, with or without tuning. Public model usage The buzz about GAI has been created by the public availability of several tools that have been built on GAI models. Some of the most popular tools at present (at the time of writing, July 2023), and the models they are based on, are shown in Table 3. In addition to their web/mobile interfaces, several of these services also provide public application programming interfaces (APIs), enabling developers to build applications that incorporate GAI Current popular generative AI tools and models Tool Type Model ChatGPT Text to text GPT-3.5/GPT-4 Bing Chat Text to text GPT-4 Google BARD Text to text PaLM 2 Anthropic Claude Text to text Claude 2 Midjourney Text to image Midjourney v5 Bing Image Generator Text to image DALL-E 2 Dream Studio Text to image Stable Diffusion GitHub Copilot Text/code to code OpenAI Codex (based on GPT-3) Applied Marketing Analytics Vol. 9, 3 270–280 © Henry Stewart Publications 2054-7544 (2023) Using generative AI to turbocharge digital marketing functionality very easily. However, there are several serious limitations with using these tools, as set out below. Data security/privacy When ChatGPT launched, information sent in prompts was used to train the models, meaning that confidential information provided as part of prompts could find its way into responses. Since March 2023, OpenAI has changed its policy — prompt data is only used for training if the user opts in. Nevertheless, users of services based on public hosted models should not include personal or company confidential information in their prompts (in much the same way that people should not include personal information in their search queries). This issue has caused many organisations to ban the use of ChatGPT altogether,9 especially in the financial sector. No specialised expertise The default version of public models, while well versed in a huge range of content from the public Internet, cannot answer questions or create content based on private information (such as a company’s product details) to which they do not have access. While it is possible to tune the performance of these public models using prompt engineering and embedding techniques, as explored below, those techniques fall foul of the limitations around data privacy and security already mentioned above. Cost The pricing models for public GAI tools are based on usage and data volumes. The OpenAI APIs are priced by ‘token’ (both on the prompts and responses). A token is roughly four characters or about 0.75 of a word. The GPT-4 API currently costs between US$0.03 and US$0.06 per thousand tokens. Feeding complex prompts to these services at high volume could thus generate quite significant costs — for example, a thousand 1,000-token prompts generating a thousand 500-token responses on GPT-4 would cost US$60 (though GPT-3 is currently around a tenth of this price). Private instance of cloud-hosted foundational model A better approach is to use one of the available hosted GAI models provided by the major cloud providers. Amazon’s Bedrock service provides an environment for hosting generative models from Anthropic, Stability.­ai (Stable Diffusion) and Amazon itself. Microsoft provides the Azure OpenAI service which makes it easy to get up and running with GPT and DALL-E models from OpenAI, while Google’s Vertex AI platform supports its own set of GAI models, such as PaLM, Imagen and Codey. The first virtue of a dedicated cloudhosted GAI model is that it provides a much more secure environment for data that may be passed through prompts. Although organisations should still pay attention to data security and privacy issues (as discussed in more detail below), these will more closely resemble the kinds of issues the organisation is already addressing in its use of cloud-hosted services. The second advantage of this approach is that it enables the organisation to customise the model to address the target use case more effectively. For example, imagine an Internet service provider (ISP) that provides home Internet service and wants to deploy an LLMpowered chatbot to help customers diagnose connectivity issues at home. The following prompt entered into ChatGPT generates the answer below (summarised by ChatGPT from a longer, more comprehensive answer): PROMPT: My internet at home isn’t working. How can I fix it? © Henry Stewart Publications 2054-7544 (2023) Vol. 9, 3 270–280 Applied Marketing Analytics 275 Thomas ANSWER (summarized): 1. Check modem and router connections. 2. Power cycle modem and router. 3. Restart your devices. 4. Verify signal lights on modem and router. 5. Test a wired connection to isolate Wi-Fi issues. 6. Reset your router to default settings. 7. Check for service outages with your ISP. 8. Update firmware and drivers for modem, router, and devices. 9. Contact your ISP’s technical support for further assistance. This is a good and helpful set of generic suggestions, but if the ISP wants to include any information that is specific to their businesses or products (such as how to use the web to check for service outages, or the connection from the router) they need to customise the language model with this information. Broadly, there are two ways of doing this: 1. prompt engineering: structuring and augmenting prompts to create more helpful responses; 2. model tuning: augmenting the foundational model with additional information. Prompt engineering Prompt engineering is the process of including structured information in the prompts provided to text-input-based GAI services to guide their responses. Including the right additional information within prompts can significantly improve the quality of responses. Some key prompt engineering techniques include: Creating the right system message: The system message is a special message to the LLM that sets its overall behaviour. For example: ‘You are a helpful assistant working for XYZ corp, an ISP. You can answer questions about internet connectivity 276 and account management. If you are asked other questions, answer, “I’m afraid that’s outside my expertise area.”’ Use few-shot learning: You can increase the chance that an LLM will generate a useful response by providing several examples of what a successful response looks like before asking your question. Provide clear instructions and prime the output: By providing clear instructions for the LLM (for example, ‘Provide a step-by-step guide to diagnosing Internet connectivity problems’ and repeating them at the end of the prompt (eg ‘Step 1’:) you can ensure the response will be effective and in the format that you need. Use chain-of-thought prompting: LLMs will often generate better results if asked to include their train of thought within the answer. Integrate external data: An app built on top of an LLM can call out to an external data source and include data from this source in the prompt. This could be search engine results, or relevant information about the user entering the query. In our ISP example, this could be information about the user’s hardware setup. For working with large quantities of external data, embeddings can be used. Embeddings OpenAI’s GPT-4 model API will accept prompt input of up to 32,768 tokens (about 24,500 words) — equivalent to a 100-page novel, while Anthropic’s Claude model will handle up to 100,000 tokens. This means that entire documents can be passed through prompts and the LLM asked to perform tasks such as summarisation, translation or style modification or to use the content as context to answer specific questions. In our example above, the ISP could pass their customer service manual through the prompt to provide the context to answer the question. However, passing a large block of content through every prompt session will cause costs to be high — for example, at current Applied Marketing Analytics Vol. 9, 3 270–280 © Henry Stewart Publications 2054-7544 (2023) Using generative AI to turbocharge digital marketing prices,10 a single 24,000-word prompt to GPT-4 that generates a 1,000-word response (say, a summary of the content in the prompt) would cost around US$2. It would not take very long for the cost of such queries to mount up. A way to address this issue and enable prompts to incorporate context from even larger corpuses (including image data) is to use embeddings. Embedding is the process of creating a simplified multidimensional representation (a vector) of input data (such as words, sections or text, or regions of an image). Input elements with vectors that are close together in multidimensional space are more similar than those that are far apart. Embeddings have been used for some time for natural language processing, having been proposed by Google researchers in a paper in 2013.11 They can be used to augment prompt context via a three-step process: • Step 1: Pass the supporting content through an embeddings API to generate a vectorised index into the content and store this. • Step 2: When a user enters a prompt, vectorise the prompt and use this vectored representation to retrieve relevant sections of the supporting material. • Step 3: Append the retrieved supporting material to the user’s prompt and pass to the generative AI model. Cloud vendors are already creating services to make this process easier. Microsoft Azure’s ‘OpenAI on your data’ offering12 uses the Azure Semantic Search service (which uses embeddings to vectorise words and text) in conjunction with the OpenAI GPT APIs to enable just this kind of use case. An easy-to-use consumer-facing example of this kind of setup is the tool ChatPDF, which enables the user to upload a PDF file and then use a chat interface to ask questions about it. Embeddings are also used in some image generation models, such as Stable Diffusion, to enable a starting set of images to be used as the baseline for the model so that it generates output with a more consistent style.13 Model tuning Effective use of prompt engineering and embeddings can improve the performance of a GAI model, but for some scenarios it will be necessary to fine-tune the model itself. Examples of these scenarios include: • ensuring the best performance across a defined range of inputs on a specialised topic; • tuning the model to a specific kind of task, such as sentiment analysis. Model tuning is a more time-intensive process than prompt engineering and involves creating a structured set of data in a prompt/response format, where each prompt/response pair provides a training example for the model. For example, for our ISP, an example prompt/response pair might be: Prompt Response How do I check if my Netgear Nighthawk R7000 router is connected to the Internet? First check that the LED on/off switch on the back of the router is switched to the ON position. The Internet LED on the top panel of the router (second from the left) will be illuminated with a solid or blinking white light if the router is connected to the Internet. If the light is solid amber, the router is connected to the modem but no connection to the Internet is detected. © Henry Stewart Publications 2054-7544 (2023) Vol. 9, 3 270–280 Applied Marketing Analytics 277 Thomas For most training scenarios, several hundred such prompt/response pairs will be needed. The benefits of model tuning over prompt engineering are as follows: • Prompts to a tuned model can be shorter as they do not need to include contextual data (since this is already now in the model), saving token costs. • Because of the shorter prompts, the model will return answers more quickly. On the other hand, tuning models comes with some downsides: • Creating the training material is timeconsuming. • A tuned model is likely to be more expensive to use per token (for example, a tuned version of GPT-3 is about a hundred times more expensive per token than its untuned counterpart). • There is certain data that you cannot use to tune a model (for example, quickly changing data or customer data). Running a private model on-premise on private cloud Many use cases for GAI will rely on data that must be very tightly controlled and perhaps restricted to on-premise or private cloud environments. In this case an organisation will need to provide the hardware to train (as necessary) and run the model itself. Many open-source models (including versions of GPT and Stable Diffusion) are available on Hugging Face.14 NVIDIA’s Triton Inference Server provides a platform for training and deploying models on a local server, taking advantage of multiple CPUs and GPUs for model scaling. A benefit of using models in this way (whether on-premise or in the cloud) is that the cost profile is determined by the compute and storage required to support them, rather than a token-based cost. This 278 can have benefits in a more scaled environment, especially if the model needs to be tuned. IMPLEMENTATION CONSIDERATIONS AI regulation Article 22 of the General Data Protection Regulation (GDPR) grants data subjects the right not to be subject to automated decisions (eg by AI systems) that have a significant or legal impact on their lives (such as being offered a loan or a job). Most automated decision making for marketing (eg advert targeting) would be unlikely to meet this criterion, though dynamic pricing, or making someone aware of, say, an insurance product that another person was not made aware of, could potentially reach this bar. Relatively few GDPR enforcement cases have listed a breach of Article 22 as a factor, but the Italian DPA fined two food delivery companies — Deliveroo15 and Foodinho16 — a combined total of €5.1m for their use of AI models to assess their drivers’ performance in a way that was neither transparent nor deemed necessary for the functioning of their businesses. In addition to the GDPR, the EU is (at the time of writing) proposing a draft law to regulate the use of AI, known as the EU AI Act.17 The final form of the law is still being worked out with the member states, but the core framework will be based on a classification of AI systems according to the risk they pose, in three categories: • unacceptable risk (banned); • high-risk (allowed but must be registered and conformity with regulation established); • low-risk (no action required). Generative AI is mentioned specifically in the draft legislation, primarily in the context of notification to users of its use. Organisations will be required to: Applied Marketing Analytics Vol. 9, 3 270–280 © Henry Stewart Publications 2054-7544 (2023) Using generative AI to turbocharge digital marketing • disclose that content was generated by AI; • design the model to prevent it from generating illegal content; • publish summaries of copyrighted data used for training. These requirements provide an incentive for organisations to use models which can provide information about the copyrighted data they have used for training and may even result in EU-specific versions of public models. Privacy and GDPR Organisations that have used third party services (such as public cloud platforms or SaaS products like Salesforce) have long had to ensure that these services (and their providers) are GDPR-compliant; as data processors under the law, they have a responsibility to store and process the data safely, retain it for an appropriate period and facilitate the retrieval or deletion of the data in response to a data subject request. The foundational GAI models provided by the likes of Microsoft and Google are available as hosted services on those vendor’s public cloud platforms, and so the same rules will apply to the ingestion of any personal data into them, most notably that personal data can be removed from them at a data subject’s request. For this reason, personal data must not be used in fine-tuning models, as it will be impossible to remove subsequently without entirely retuning the model, which will be resource-intensive and expensive. Organisations using cloud-hosted models will need to pay attention to cross-border data transfer, especially transfer of data out of the EU. The major cloud vendors are making their GAI models available in different regions around the world, but organisations should pay attention to where apps built on top of such services are hosting their data. In the rush to get into the GAI market, some app vendors may take shortcuts on areas like privacy policies and practices, so these should be carefully assessed. Non-AI regulation Because GAI models are generating content and supporting automated decision making, their outputs will be subject to other regulation, such as advertising standards and copyright law. Creating many variants on advertising creative using a GAI model, especially using contextual or customer data, increases the risk that one of the variants may be misleading (for example, offering a service in an area where it is not available or conflating two conflicting product attributes). Advertising regulatory bodies like the UK’s Advertising Standards Authority are already strengthening their own use of machine learning to be able to detect misleading or offensive adverts at scale. Complying with appropriate copyright laws is another challenge that GAI-created content will need to contend with. OpenAI and Meta are the subject of multiple lawsuits18–20 claiming that they have used copyrighted work to build their models; organisations using GAI for content creation could find themselves caught up in these lawsuits or new ones if the generated content too closely resembles copyrighted works. The risk of accidental (or lazy) plagiarism using LLMs is high. SUMMARY Generative AI will have a significant impact on the digital marketing industry, particularly in the area of automated creative generation, though its applications will deliver wider benefits around integration and analytics. Organisations using GAI tools should pay close attention to data security, privacy and ethical considerations to ensure that they do not create reputational or legal risk through the use of this new technology. © Henry Stewart Publications 2054-7544 (2023) Vol. 9, 3 270–280 Applied Marketing Analytics 279 Thomas References 1. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L. and Polosukhin, I. (2017) ‘Attention Is All You Need’, arXiv:1706.03762, available at https:/­/­arxiv.­org/­abs/ ­1706.­03762 (accessed 8th November, 2023). 2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A. and Bengio, Y. (2014) ‘Generative Adversarial Networks’, arXiv:1406.2661, available at https:/­/­arxiv.­org/­abs/ ­1406.­2661 (accessed 6th October, 2023). 3. Brown, T. et al. (2020) ‘Language Models Are Few-Shot Learners’, arXiv:2005.14165, available at https:/­/­arxiv.­org/­abs/­2005.­14165 (accessed 6th October, 2023). 4. Albergotti, R. (2023) ‘The Secret History of Elon Musk, Sam Altman, and OpenAI’, available at https:/­/­ www.­semafor.­com/­article/­03/­24/­2023/­the-­secret- ­history-­of-­elon-­musk-­sam-­altman-­and-­openai (accessed 14th July, 2023). 5. Okudaira, K. (2023) ‘Meta to Debut Ad-creating Generative AI this Year, CTO Says’, available at https:/­/­asia.­nikkei.­com/­Business/­Technology/ ­Meta-­to-­debut-­ad-­creating-­generative-­AI-­this-­year- ­CTO-­says (accessed 26th September, 2023). 6. Dave, P. (2023) ‘Google Will Soon Show You AI-Generated Ads’, available at https:/­/­www.­wired. ­com/­story/­google-­chatgpt-­ai-­generated-­ads/­ (accessed 26th September, 2023). 7. Foster, W. (2023) ‘Putting AI and Automation to Work for You: An Open Letter from Zapier’, available at https:/­/­zapier.­com/­blog/­ai-­open-­letter/­ (accessed 26th September, 2023). 8. Thomas, I. (2023) ‘The Great SQL Bot Bake Off: Comparing the Big LLM Beasts on SQL Code Generation’, available at https:/­/­www.­liesdamnedlies. ­com/­2023/­07/­the-­great-­sql-­bot-­bake-­off- ­comparing-­the-­big-­llm-­beasts-­on-­sql-­code- ­generation.­html (accessed 26th September, 2023). 9. O’Brient, S. (2023) ‘7 Companies That Have Banned Employees from Using ChatGPT’, available at https:/­/­investorplace.­com/­2023/­05/­7-­companies- ­that-­have-­banned-­employees-­from-­using-­chatgpt/­ (accessed 26th September, 2023). 10. OpenAI (n.d.) ‘Pricing’, available at https:/­/­openai. ­com/­pricing (accessed 26th September, 2023). 280 11. Mikolov, T., Chen, K., Corrado, G. and Dean, J. (2013) ‘Efficient Estimation of Word Representations in Vector Space’, arXiv:1301.3781, available at https:/­/­ arxiv.­org/­abs/­1301.­3781 (accessed 6th October, 2023). 12. Microsoft (n.d.) ‘Azure OpenAI on Your Data (preview)’, available at https:/­/­learn.­microsoft.­com/ ­en-­us/­azure/­ai-­services/­openai/­concepts/­use-­your- ­data (accessed 26th September, 2023). 13. Skelton, J. (2022) ‘Stable Diffusion Tutorial Part 2: Using Textual Inversion Embeddings to Gain Substantial Control over Your Generated Images’, available at https:/­/­blog.­paperspace.­com/ ­dreambooth-­stable-­diffusion-­tutorial-­part-­2-­textual- ­inversion/­(accessed 26th September, 2023). 14. Hugging Face (n.d.) ‘Home Page’, available at https://huggingface.co/. 15. GPDP (2021) ‘Ordinanza ingiunzione nei confronti di Deliveroo Italy s.r.l.’, available at https:/­/­www. ­gpdp.­it/­web/­guest/­home/­docweb/­-­/­docweb- ­display/­docweb/­9685994 (accessed 26th September, 2023). 16. GPDP (2021) ‘Ordinanza ingiunzione nei confronti di Foodinho s.r.l.’, available at https:/­/­www.­gpdp.­it/ ­web/­guest/­home/­docweb/­-­/­docweb-­display/ ­docweb/­9675440 (accessed 26th September, 2023). 17. European Parliament (2023) ‘EU AI Act: First Regulation on Artificial Intelligence’, available at https:/­/­www.­europarl.­europa.­eu/­news/­en/ ­headlines/­society/­20230601STO93804/­eu-­ai-­act- ­first-­regulation-­on-­artificial-­intelligence (accessed 26th September, 2023). 18. LLM Litigation (2023) ‘Case 3:23-cv-03223, Tremblay/Awad vs OpenAI Inc.’, available at https:/­/ ­llmlitigation.­com/­pdf/­03223/­tremblay-­openai- ­complaint.­pdf (accessed 26th September, 2023). 19. LLM Litigation (2023) ‘Case 3:23-cv-03416, Silverman/Golden/Kadrey vs OpenAI Inc.’, available at https:/­/­llmlitigation.­com/­pdf/­03416/­silverman- ­openai-­complaint.­pdf (accessed 26th September, 2023). 20. LLM Litigation (2023) ‘Case 3:23-cv-03417, Kadrey/Silverman/Golden vs Meta Platforms Inc.’, available at https:/­/­llmlitigation.­com/­pdf/­03417/ ­kadrey-­meta-­complaint.­pdf (accessed 26th September, 2023). Applied Marketing Analytics Vol. 9, 3 270–280 © Henry Stewart Publications 2054-7544 (2023)