Judicial Decision under Ambiguity and Predictive Justice Sébastien Massoni1 1 BETA Vincent Teixeira1 - Université de Lorraine L2 - Défis Economiques du XXIe siècle March 23, 2023 Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 1 / 42 Predictive Justice Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 2 / 42 Predictive Justice COMPASS algorithm was used in Wisconsin to estimate risk assessment in sentencing despite a non-disclosed algorithm, biased toward minority, not developed for sentencing and predicting at the group level (State v. Loomis, Kehl et al., 2017). Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 2 / 42 Predictive Justice Natural Language Processing and Machine Learning algorithms reach 79% of accuracy on predicting the outcome of cases tried by the European Court of Human Rights based solely on textual content (Aletras et al., 2016). Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 2 / 42 Predictive Justice A.I. ’Prosecutor’ that can press charges with more than 97% accuracy in China. It can identify and press charges for Shanghai’s 8 most common crimes (e.g. credit card fraud, running a gambling operation, etc.). Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 2 / 42 LegalTechs and Predictive Justice Definitions LegalTech: Use of technology and software to provide legal services and support the legal industry; often associated with technology startups disrupting the practice of law. Predictive justice: computer tool used to carry out statistical analyses based on a big data extracted from case law, i.e. court decisions previously rendered. Predictive justice algorithms might: Process behavioral data to assess individual behaviors (e.g. risk of recidivism); Analyze legal data to produce decision models and anticipate court decisions (similar cases, effective legal arguments, assessment of chances, average amount perceived or procedure duration). Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 3 / 42 Characteristics of the Law Market Credence goods market with law agents (lawyers, judges, etc.) and litigants. Before digitalization After digitalization Regulated profession Break of the oligopoly Noisy information (reputation, advertisement) More objective information Transaction costs Potential decrease of costs Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 4 / 42 LegalTechs and Information LegalTechs brought 3 types of innovation: Intermediation - relation between justice and its actors; Processing and provision of data; Artificial Intelligence (A.I). The two first types of innovation have a common point: they bring new information. We design two studies to understand how this new source of information might change behaviors in ambiguous judicial decisions. We use real judicial cases and a predictive justice tool: Case Law Analytics. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 5 / 42 Case Law Analytics Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 6 / 42 Case Law Analytics Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 6 / 42 Case Law Analytics Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 6 / 42 Case Law Analytics Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 6 / 42 Case Law Analytics Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 6 / 42 Risk and Ambiguity Risk: Known probabilities Ambiguity: Unknown probabilities Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 7 / 42 Risk and Ambiguity Risk: Known probabilities Ambiguity: Unknown probabilities Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 7 / 42 About Ambiguity It requires new models: e.g. multiple priors, probability weighting. Two main measures of ambiguity attitudes (Abdellaoui et al., 2011): Ambiguity aversion / pessimism; Likelihood insensitivity / perceived level of ambiguity. Ambiguity aversion is generally observed but not universal (e.g. Kocher et al., 2018, for losses and lower likelihoods). Two questions related to our study: Sources of ambiguity; Value of information. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 8 / 42 Sources of Ambiguity: Artificial and Natural Events Sources of ambiguity matter (Tversky & Fox, 1995; Abdellaoui et al., 2011). Artificial events (e.g. lotteries and urns) Urns allow for a control of likelihood beliefs; But their external validity is questioned (Heath & Tversky, 1991; Camerer & Weber, 1992). Natural events (e.g. weather and stock markets predictions) Difficulties to estimate likelihood beliefs may prevent the identification of ambiguity attitudes; Ambiguity attitudes are closer to rationality (Li et al., 2018). Measurements without likelihood beliefs (Baillon et al., 2018) Ambiguity aversion and ambiguity generated insensitivity (perception of ambiguity); Indexes are valid for artificial or natural events and under many ambiguity theories. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 9 / 42 Information in Decision under Ambiguity Value of information Ambiguity aversion increases the value of information (Snow, 2010; Attanasi & Montesano, 2012; Kops & Pasichnichenko, 2020); But under-valuation of information not resolving ambiguity (Ambuehl & Li, 2018; Li, 2020). Beliefs updating and ambiguous information Failure of Bayesian updating (Epstein & Halevy, 2019); Insensitivity or ambiguity aversion in information accuracy (Liang, 2021; Shishkin & Ortoleva, 2021). Consequentialism and dynamic consistency Ambiguity aversion increases violation of dynamic consistency but not consequentialism (Dominiak et al., 2012; Bleichrodt et al., 2018; Galanis, 2021). Information avoidance is common for real-life decisions Health, finance, managerial decision (see Loewenstein’s works). Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 10 / 42 Judicial Decision and Natural Ambiguity Judicial decisions offer a natural source of ambiguity. We take real cases that are already judged by a court. We then have a situation without artificial ambiguity. We can provide objective information through predictive justice. We have different situations: complete ambiguity, partial ambiguity, risk and similarity. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 11 / 42 Study 1: Ambiguity Attitudes for Judicial Events We want to compare behaviors under artificial and natural events (within-subject). For three additional types of information (between-subject): [0; 100]: C-Ambiguity (baseline); link p1 ; p2 ; p3 : Risk with probability distribution; [pi ; pi ]: P-Ambiguity with confidence index; Similarity with distribution of close cases. Massoni & Teixeira Ambiguity and Predictive Justice link link link L2 Défis Economiques 12 / 42 Types of Ambiguity Decisions under natural events are all ambiguous: risk information leads to probabilistic ambiguity due to the uncertainty of the predictive algorithm. While decisions under artificial events with risk information is a standard risky choice. Similarity can be interpreted within the case-based framework (Gilboa & Schmeidler, 1995, 2001). Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 13 / 42 Stimuli Subjects will have to take two types of decision: 3-color Ellsberg urns and 3 amounts of condemnation in a divorce situation. We elicit the certainty equivalent by probability matching for each case (cf. Baillon et al., 2018). We will measure attitudes toward ambiguity through 2 indexes: ambiguity aversion; a-insensitivity (level of ambiguity). They will all face C-Ambiguity. By providing additional information we will implement Risk, P-Ambiguity and Similarity. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 14 / 42 Matching Probability Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 15 / 42 Matching Probability Three mutually exclusive and exhaustive non-null events E1 , E2 and E3 where Ei is a single event and Eij a composite event (Ei ∪ Ej ). Matching probability mE of event E means that receiving X (here e20) under event E is equivalent to receiving X with probability mE . XE 0 ∼ XmE 0 Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 15 / 42 Matching Probability The ambiguous lottery XE 0 is evaluated by its SEU: P (E)U (X) + (1 − P (E))U (0) The risky lottery XmE 0 is evaluated by: mE U (X) + (1 − mE )U (0) For ambiguity-neutral DM: P (E) = mE : by laws of probability mE + mE C = 1. Empirically we often observe P (E) ̸= mE : size and sign of deviations from laws of probability reflect ambiguity attitudes; e.g. ambiguity averse means mE + mE C < 1. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 16 / 42 Matching Probability Relative elevation of MP function: extent of deviation toward ambiguity neutrality: ambiguity aversion (seeking) for likely (unlikely) events. Relative flatness in the middle: (in)sensitivity of the DM in discriminating different intermediate levels of likelihood of ambiguous events. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 17 / 42 Matching Probability We observe P (E) − mE . With known beliefs: mE captures ambiguity attitudes controlling for risk aversion (Dimmock et al., 2016): P − m. With unknown beliefs, Bailon et al. (2018) methods is used: Elicitation of mEi for 3 mutually exclusive and exhaustive non-null events and for their composites; Allows to estimate ambiguity attitudes irrespective of risk aversion and subjective beliefs. E1 , E2 , E3 , {E1 ; E2 }, {E2 ; E3 }, {E1 ; E3 } Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 18 / 42 Matching Probability E1 , E2 , E3 , {E1 ∪ E2 }, {E1 ∪ E3 }, {E2 ∪ E3 } Average single-event MP: ms = (mE1 + mE2 + mE3 )/3; Average composite-event MP: mc = (mE12 + mE13 + mE23 )/3; Under ambiguity neutrality: ms = Massoni & Teixeira 1 3 and mc = 23 . Ambiguity and Predictive Justice L2 Défis Economiques 19 / 42 Ambiguity aversion: b index Ambiguity aversion index is: b = 1 − (ms + mc ) with: ambiguity neutrality for b = 0; ambiguity aversion for b ∈]0; 1]; ambiguity seeking for b ∈ [−1; 0[. Ambiguity averse DM has low m i.e. is willing to pay a premium (in winning probabilities) to avoid ambiguity. Maximally ambiguity averse DM has all m = 0 and thus b = 1. The reciprocal is true for ambiguity seeking DM. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 20 / 42 A(mbiguity-generated)-insensitivity: a index Ambiguity-generated insensitivity index is: 1 a = 3 × ( − (mc − ms )) 3 with Ambiguity neutrality for a = 0 (perfect discrimination between single and composite events). Maximal insensitivity for a = 1 (i.e. mc = ms ). a < 0 corresponds to over-sensitivity, irrationality or noise. The better a DM can discriminate between single and composite events, the larger mc − ms is. Maximal insensitivity DM treats all uncertainties as fifty-fifty. Measure of perception of ambiguity: the more ambiguity a DM perceives, the more the likelihood of events are perceived as one blur. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 21 / 42 Ambiguity Attitudes indexes b index captures a motivational component on ambiguity attitudes. a index captures a cognitive component on ambiguity perception. The two indexes are orthogonal i.e. ambiguity may still play a role through insensitivity if there is no aversion. These indexes unify and extend many existing indexes without the need of specific assumptions. e.g. multiple priors: α = Massoni & Teixeira b 2a + 1 2 Ambiguity and Predictive Justice L2 Défis Economiques 22 / 42 Assumptions H1 Ambiguity aversion is affected by the nature of the events: (a) Ambiguity aversion for artificial events: bartif icial > 0 (b) Absence of ambiguity aversion for natural events: bnatural ≤ 0 (c) Difference of ambiguity aversion between natural and artificial events: bartif icial ̸= bnatural H2 Ambiguity aversion decreases with the content of information: (a) bambiguity > bothers (b) bsim > bpartial > brisk H3 A-insensitivity is affected by the nature of the events: (a) A-insensitivity for artificial events: aartif icial > 0 (b) A-insensitivity for natural events: anatural > 0 (c) Higher a-insensitivity for natural events: anatural > aartif icial H4 A-insensitivity decreases with the content of information: (a) aambiguity > aothers (b) asim > apartial > arisk Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 23 / 42 Study 1 Experiment took place at the Laboratoire d’Economie Experimentale de Strasbourg (LEES) in November 2021. There were 252 subjects (gender balanced) for a total of 12 sessions (4 by treatment): 81 subjects for Risk; 88 subjects for P-Ambiguity; 83 subjects for Similarity. The average payoff was e17 for an average time of 1h30. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 24 / 42 Timeline of decisions Part 1: complete ambiguity: 12 rounds (case 1: 1-6; case 2: 7-12). Part 2: one source of information (risk, partial, similarity): 12 rounds (same cases). Two parts with natural events and two parts with artificial events (pseudo-randomized for order effect). Check for consistency: no differences for indexes between rounds 1-6 and 7-12 within a same types of ambiguity. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 25 / 42 Results: b index Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 26 / 42 Results H1 Ambiguity aversion is affected by the nature of the events: YES Ambiguity aversion for artificial events: bartif icial > 0 YES Absence of ambiguity aversion for natural events: bnatural ≤ 0 YES Difference of ambiguity aversion between natural and artificial events: bartif icial ̸= bnatural H2 Ambiguity aversion decreases with the content of information: YES in artificial bambiguity > bothers YES only for risk in natural bambiguity > bothers NO bsim > bpartial > brisk Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 27 / 42 Results: a index Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 28 / 42 Results H3 A-insensitivity is affected by the nature of the events: YES A-insensitivity for artificial events: aartif icial > 0 YES A-insensitivity for natural events: anatural > 0 NO Higher a-insensitivity for natural events: anatural > aartif icial H4 A-insensitivity decreases with the content of information: YES in artificial aambiguity > aothers YES in natural except similarity aambiguity > aothers NO asim > apartial > arisk Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 29 / 42 Study 2: Value of Information We want to measure how much subjects value each type of information. Subjects face C-Ambiguity and give their matching probability. Then, they will give they Willingness To Pay (WTP) for a source of information. BDM mechanism will determine their access to the information. Subjects will face each type of information for both sources (artificial and natural). Due to our previous results, subjects will always face natural events first. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 30 / 42 Types of Ambiguity Subjects play the different types of ambiguity without knowing the output of the BDM. Except for WTP = 0, they are facing a double ambiguous events: on the probability and on the outcome. See multidimensional ambiguity (Eicheberger et al., 2015; Eliaz & Ortoleva, 2016; Aggarwal & Mohanty, 2021). This aspect is not in the scope of the current studies. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 31 / 42 Assumptions H5 Under-evaluation of information not resolving ambiguity: W T Partif icial > W T Pnatural H6 Increase of valuation with the content of information: W T P risk > W T P partial > W T P sim H7 Increase of valuation with the level of ambiguity aversion: ρ(b, W T P ) > 0 H8 Link between the level of sensitivity to ambiguity and the valuation: ρ(a, W T P ) ̸= 0 Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 32 / 42 Study 2 Experiment took place at the Laboratoire d’Economie Experimentale de Strasbourg (LEES) in January 2021. There were 247 subjects (gender balanced) and a total of 12 sessions (4 by treatment): 81 subjects played Risk first; 83 subjects played P-Ambiguity first; 83 subjects played Similarity first. The average payoff was 17 euros for an average time of 1h45. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 33 / 42 Timeline of decisions Part 1: complete ambiguity (6 rounds), WTP for information 1, information 1 (6 rounds same case). Part 2: complete ambiguity (6 rounds new case), WTP for information 2, information 2 (6 rounds same case). Part 3: complete ambiguity (6 rounds new case), WTP for information 3, information 3 (6 rounds same case). Order of information’s type pseudo-randomized for order effect. 3 parts with natural events first and then 3 parts with artificial events (not-randomized). Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 34 / 42 Results: WTP Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 35 / 42 Results H5 Under-evaluation of information not resolving ambiguity: YES W T Partif icial > W T Pnatural H6 Increase of valuation with the content of information: NO W T P risk > W T P partial > W T P sim Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 36 / 42 Results: WTP and ambiguity attitudes Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 37 / 42 Results H7 Increase of valuation with the level of ambiguity aversion: YES only in artificial ρ(b, W T P ) > 0 H8 Link between the level of sensitivity to ambiguity and the valuation: YES ρ(a, W T P ) ̸= 0 i.e. ρ(a, W T P ) < 0 Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 38 / 42 Additional results The two indexes of ambiguity attitudes are theoretically orthogonal but we observe a strong correlation between a and b indexes: for artificial events in all types of ambiguity; but not for natural events except for P-ambiguity; These correlations are found in both studies. Decision time seems to be linked to lower b and a indexes: The longer a DM think about the decision, the closer he is to ambiguity neutrality. Regressions analyses confirm or reinforce main and additional results. Robustness to outliers (violation of monotonicity, irrational a, etc.). Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 39 / 42 Conclusion H1 Ambiguity aversion is affected by the nature of the events: YES Ambiguity aversion for artificial events: bartif icial > 0. YES Absence of ambiguity aversion for natural events: bnatural ≤ 0. YES Difference of ambiguity aversion between natural and artificial events: bartif icial ̸= bnatural H2 Ambiguity aversion decreases with the content of information: YES in artificial bambiguity > bothers YES only for risk in natural bambiguity > bothers NO bsim > bpartial > brisk H3 A-insensitivity is affected by the nature of the events: YES A-insensitivity for artificial events: aartif icial > 0. YES A-insensitivity for natural events: anatural > 0. NO Higher a-insensitivity for natural events: anatural > aartif icial H4 A-insensitivity decreases with the content of information: YES in artificial aambiguity > aothers YES in natural except similarity aambiguity > aothers NO asim > apartial > arisk Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 40 / 42 Conclusion H6 Under-evaluation of information not resolving ambiguity: YES W T Partif icial > W T Pnatural H7 Increase of valuation with the content of information: NO W T P risk > W T P partial > W T P sim H8 Increase of valuation with the level of ambiguity aversion: YES only in artificial ρ(b, W T P ) > 0 H9 Link between the level of sensitivity to ambiguity and the valuation: YES ρ(a, W T P ) ̸= 0 i.e. ρ(a, W T P ) < 0 Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 40 / 42 Discussion Attitudes toward ambiguity differ under natural and artificial settings. Information provided to a complete ambiguous situation change the behaviors but with no major difference according to the content of information. Values given to these information are relatively in line with attitudes. Ambiguity aversion and ambiguity generated insensitivity are related. Judicial decision is just a framework but coming experiment on the impact of predictive justice on lawyers behaviors and courts resolution. Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 41 / 42 Thank you for your attention Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 42 / 42 Complete Ambiguity Massoni & Teixeira Go back Ambiguity and Predictive Justice L2 Défis Economiques 42 / 42 Risk Go back Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 42 / 42 Partial Ambiguity Massoni & Teixeira Go back Ambiguity and Predictive Justice L2 Défis Economiques 42 / 42 Similarity Go back Massoni & Teixeira Ambiguity and Predictive Justice L2 Défis Economiques 42 / 42