Among the fantasies raised by AI and new technologies, predictive justice tools promise to eventually replace the judicial system, by deducing from legal texts the judgment to be rendered on a new case. Ongoing experiments to introduce these legal techs into the courtroom show that it is not so simple, and that magistrates welcome them by confronting them with their own professional ethics.

The world of justice is not immune to AI: evidenced by the proliferation of legal tech , these devices based on access to large judicial databases and on more or less predictive algorithmic processing. In general, the developers’ discourse emphasizes the “disruptive” nature of their programs, as if we were wiping the slate clean of all other equipment in judicial activity.

Moreover, presenting the application of AI to the judicial field by insisting as much as the promoters of these tools do on their predictive capacities awakens the fantasy of a complete substitution of human professionals by artificial intelligence. This amounts to claiming that AI-based tools would be capable of determining judicial decisions by themselves.

Legal tech systems rely on large judicial databases and are likely to combine three types of tools: a multi-criteria search engine, statistical processing, and probabilistic and predictive algorithms for a possible decision. Each of these levels of the system is likely to help with the judicial decision, but in a different way. If the last level is the most predictive, it is the search engine, which, as we will see, will play a crucial role, even if it arouses fewer technological fantasies.

An experiment with predictive justice tools in Strasbourg

We participated in two experiments at the Strasbourg court of an IT tool exploiting large judicial databases, combining search engines and statistical processing. These experiments were carried out in collaboration with the magistrates of the Civil Division (litigation of compensation for personal injury) and the Family Affairs Division (issue of compensatory benefits in divorce cases). These experiments constitute a meeting point between the judicial world and the world of legal tech IT . They shed light on the conditions in which this type of tool, and more particularly search engines allowing “similar cases” to be brought to the attention of judges, equip, assist and transform judicial decisions.

First of all, whatever the decision-making support system envisaged (simple scale, search engine or predictive algorithm), we cannot understand this type of experimentation without referring to the ethos of magistrates. Ethos intersects with the dispositions and values ​​internalized by them in the context of their professional activities. It determines the encounter with algorithmic tools and their promoters, and the way in which magistrates are likely to use them. The latter in fact refer to the ” imperium of the judge”, which requires that the magistrate be solely in charge of the case, fully responsible for its processing, and that his decision be written independently. These normative expectations constrain a magistrate’s ability to ask for help, and even weigh on the possibility of discussing a case with his colleagues.

Conversely, the world of IT is permeated by an ideology of all-out collaboration. Legal tech innovators sometimes fail to get magistrates to adopt their way of collaborating, which can slow down the adoption of the tools they develop and lead to the failure of certain experiments .

Identify similar cases

Our work shows that the use of these devices by magistrates is strongly constrained by the pressure to produce decisions and manage the flow of cases that weighs on them, and which greatly limits their learning opportunities. For the magistrates who used them, their use focused mainly on cases that they considered particularly complex or atypical (complexity of professional situations, assets or retirement rights for example for the disputes studied here). These tools allowed them to conduct a search in the database of similar cases, where previous decisions were likely to shed light on this complexity of the present cases. On the one hand, for these potentially long-to-process cases, the cost and learning time of the new tool seemed more acceptable. On the other hand, for delicate cases, having similar cases could prove enlightening.

It should be noted that, in addition to the different nature of the notion of “proximity” between different cases, magistrates have had to develop new skills in handling search criteria, in order to be able to isolate neither too many nor too few similar cases.

It is in the relationship to these similar cases that the role of the experimental devices, used almost exclusively as search engines by magistrates, was played out, even if they also allowed statistical processing. However, even this more familiar use of digital technology poses ethical problems for magistrates, when it allows them to find cases that are close enough to guide the decision. At one extreme, if the judge treats the cases and relevant decisions returned by the search engine as additional information, but keeps them at a distance, the technology is only one tool among others, such as the various paper charts or scales that they may already have. At the other extreme, if the magistrate treats a related case as a source of law (i.e. a precedent that sets a precedent), and “models” his decision on it, then it is as if the technology largely produced the decision. And our investigation shows how sensitive this is a question, since French law excludes the logic of precedent, unlike Anglo-Saxon law.

The risk of precedent

The case of a magistrate of the Civil Division illustrates this point well. He used the search engine for a complex case: “I found a situation, which happened overseas, quite similar… I looked through them all, but this is the one that best suited my case”. At first, he specifies that he was only “inspired” by the decision: “So, I did not appropriate it, I was inspired by it…”. This nuance appears defensive, it aims to avoid giving the impression that the judge treated the case as a binding precedent (which would amount to treating the case as a source of law). The justifications he provides immediately afterwards show this: “It is because in France, unlike […] other case law systems in other countries, we do not have a culture of precedent… We have a legal system, case law being a source, one of the sources, but we must not issue what are called regulatory judgments”.

“To appropriate” the case would be to take the motivation of the decision as it is: “To appropriate it would have been, in fact, to say that I take exactly the motivation, and I reproduce it as it is, and there, I would have created a precedent”. We cannot refer to the previous case either, we must decide anew each time: “We cannot say, reference case law such, we apply exactly the same thing as the previous case law, we must, case by case, decide”. It is also necessary that the fact that we actually decide on the case be made explicitly manifest in the decision itself: “so, my reasoning, precisely in order not to take the rule of precedent, and in order not to purely apply case law, I always recall, at the beginning of each item of damage, the rule of law”.

Here we can clearly see how the combination of large labelled judicial databases and search engines increases the risk of finding similar cases and treating them as precedents, and how this temptation should be resisted. It is in this tension and the way in which it is negotiated that the role of these new judicial technologies is played out.

Finally, one of the important conclusions of our experiments is that this tension is more or less acute depending on the type of litigation envisaged. The temptation of precedent is all the stronger when the litigation is codifiable and formalizable (as is the case, for example, with compensation for personal injury). On the other hand, it is weaker for a type of litigation such as compensatory benefits in divorce cases where the cases are more singular and also involve extra-legal assessment elements. More generally, we will remember that, contrary to the ambient discourse which envisages the application of AI in justice in general, it is necessary to use a finer grain, and to consider its potential effects on a case-by-case basis, litigation by litigation.

Author Bio: Christian Licoppe is Professor of Sociology at Télécom Paris – Institut Mines-Télécom