In December 2025 the European Commission for the Efficiency of Justice (CEPEJ) of the Council of Europe adopted its guidelines on the use of generative artificial intelligence for courts, and included in them a figure that deserves to be read before any consideration of principle: according to a survey conducted in October 2025 among the members of the European Cyberjustice Network, 46 per cent of respondents confirmed that generative artificial intelligence is already in use in courts (CEPEJ). The task is therefore not to envisage a future scenario, but to establish the conditions under which tools already present in judicial offices may continue to be used there.
The high-risk category of Annex III, point 8: what it covers and what it does not
Regulation (EU) 2024/1689 does not prohibit the use of artificial intelligence in judicial activity. It classifies it as high-risk, which means that it permits it under stringent conditions. Annex III, point 8(a), includes among high-risk systems those “intended to be used by a judicial authority or on their behalf to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts”, as well as those used in a similar way in alternative dispute resolution (European Commission).
Two clarifications avoid the most frequent errors. The first concerns the breadth of the formula: the phrase “or on their behalf” brings within the perimeter the providers and consultants operating for the judicial office as well, not only direct use by the judge. The second concerns the exclusion of ancillary administrative activities, which many attribute to the Annex: that exclusion is not in Annex III, it is in recital 61, and recitals guide interpretation without having autonomous prescriptive force. Anyone building a conformity assessment on the distinction between decision-making activity and ancillary activity must be aware that they are working on an interpretative criterion, not on a textual derogation.
It should be recalled, finally, that the obligations of Chapter III for Annex III systems will apply from 2 December 2027, under the timetable redrawn by the Digital Omnibus, Regulation (EU) 2026/1744 of 8 July 2026. The Commission’s guidelines on the classification of high-risk systems, which Article 6 required to be adopted by 2 February 2026, are to date available only in draft form (European Commission).
Recital 61: support for decision-making power, not replacement
Recital 61 explains why those systems were placed among high-risk systems, and it does so in words worth reporting in full: “considering their potentially significant impact on democracy, the rule of law, individual freedoms as well as the right to an effective remedy and to a fair trial”. Immediately afterwards the European legislature fixes the point of balance: “The use of AI tools can support the decision-making power of judges or judicial independence, but should not replace it: the final decision-making must remain a human-driven activity” (European Commission).
This approach is mirrored, on the supervisory side, by Article 74(8), which for systems used in justice and democracy entrusts market surveillance to the data protection supervisory authorities or to another authority designated under the same conditions, with an explicit safeguard clause: “Market surveillance activities shall in no way affect the independence of judicial authorities or otherwise interfere with their activities when acting in their judicial capacity”. It is a delicate balance: surveillance reaches the system, not the judge.
The CEPEJ guidelines of December 2025: four principles and a rule on reasoning
The document adopted by the CEPEJ at its forty-fifth plenary meeting on 4 and 5 December 2025 rests on four statements: the exercise of judicial power is the exclusive responsibility of the courts, effective access to a human judge is always guaranteed, outputs produced by generative artificial intelligence are never binding, and the use of these tools in the drafting of decisions must be transparent.
The most significant passage, and the least cited, is however paragraph 58, which establishes that judges shall not be required to specifically justify their departure from proposals or recommendations generated by the system. The rule is less technical than it seems. If the judge had to justify the divergence, the suggestion would cease to be a suggestion and would become the default position, the one from which one departs at one’s own risk. The freedom not to give reasons for departing from the system’s suggestion is what prevents the suggestion from turning into the rule.
The guidelines build on the European Ethical Charter on the Use of Artificial Intelligence in Judicial Systems and their Environment, adopted by the CEPEJ in December 2018, which had already laid down the five principles of respect for fundamental rights, non-discrimination, quality and security, transparency, impartiality and fairness, and the principle “under user control” (Council of Europe). On the treaty level, the Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, adopted on 17 May 2024 and opened for signature in Vilnius on 5 September 2024, was ratified by the European Union on 15 May 2026 but is not yet in force, the quorum required by its Article 30 not having been reached (Council of Europe).
Article 15 of Law No 132/2025: a reservation in four paragraphs
Law No 132 of 23 September 2025, in force since 10 October 2025, devotes to the subject an article that is habitually cited for its first paragraph alone, which provides that in judicial activity “every decision on the interpretation and application of the law, on the assessment of facts and evidence and on the adoption of measures is always reserved to the judge” (Official Gazette (Gazzetta Ufficiale)). The article, however, has four paragraphs, and the other three are not without significance.
Paragraph 2 entrusts the Ministry of Justice with the task of regulating uses for the organisation of services, for the simplification of judicial work and for ancillary administrative activities: the same expression as in recital 61, used here not to exclude from the high-risk category but to found an internal regulatory power. Paragraph 3 provides that, until the European regulation has been fully implemented, experimentation and use in ordinary judicial offices are authorised by the Ministry, after consulting the national authorities. Paragraph 4 commits the Minister to promoting educational activities on artificial intelligence in the training of judges and administrative staff. To these is added Article 17, which places cases concerning the functioning of an artificial intelligence system among the matters within the exclusive jurisdiction of the tribunal under Article 9, second paragraph, of the Code of Civil Procedure.
A terminological caveat. The word “predictive” never appears in the law: the formulas “reservation of human adjudication” and “prohibition of predictive justice” are scholarly glosses, useful for expository purposes but not to be attributed to the text.
A figure to be corrected: 79 per cent, not 94
For years the Italian debate has circulated the news of a system capable of predicting the decisions of the European Court of Human Rights with 94 per cent accuracy, higher than that of the judges. The figure is inaccurate. The study referred to is that of Aletras, Tsarapatsanis, Preoţiuc-Pietro and Lampos, published in PeerJ Computer Science in 2016, and it reports an average accuracy of 79 per cent (PeerJ). The 94 per cent belongs to something else entirely: a commercial experiment of 2018 on the review of five non-disclosure agreements.
More than the number, what counts are the conditions under which it was obtained. The sample consisted of 584 cases, in English, limited to Articles 3, 6 and 8 of the Convention, with classes balanced by construction and a baseline of 50 per cent. Above all, the system did not read the parties’ applications but the judgments already drafted by the Court, a circumstance the authors openly acknowledge when they observe that the subsection on the circumstances of the case “has been formulated by the Court itself” and cannot therefore be taken as a neutral mirror of the events. Subsequent studies have further scaled down the picture: a 2020 study reports an average accuracy of 75 per cent, which falls to a range between 58 and 68 per cent when future cases are predicted on the basis of past cases, and notes that a result of 65 per cent is obtained using only the surnames of the judges on the panel (Artificial Intelligence and Law).
We consider this last figure to be the most instructive of all. A model that correctly predicts the outcome in two cases out of three knowing only the composition of the panel does not understand the law: it measures a sociological regularity. This is precisely the reason why legal reasoning cannot be reduced to calculation.
Critical aspects: human oversight is not a signature
The regulatory system described makes its own effectiveness depend on a single element: the person who decides. Article 14(4)(b) of the AI Act asks those exercising human oversight to remain aware “of the possible tendency of automatically relying or over-relying on the output produced by a high-risk AI system”, which the Italian text renders as “distorsione dell’automazione”, that is, automation bias. It is an awareness that the legislative provision alone is unable to guarantee.
The concrete risk is not the replacement of the judge, but the judge’s alignment with the system’s suggestion. If a system suggests and divergence costs time, reasoning and the possibility of censure, the reservation of decision turns into a formality: the measure is signed by the judge, but the decision is in fact determined by the system. That is why paragraph 58 of the CEPEJ guidelines, which exempts the judge from giving reasons for departing from the suggestion, carries more weight than many declarations of principle, and why paragraph 4 of Article 15 on the training of judges is not ancillary in character. Without a minimum competence in what the system does and in the data on which it was trained, human oversight is a formal fulfilment, not a control: on this point the considerations already set out with regard to the rewriting of Article 4 of the AI Act on literacy apply.
There remains the problem that our legal system addresses last: the quality of the data. A system trained on existing case law reproduces the regularities of that case law, including those no one would wish to preserve, and the individual judge is not in a position to notice, because he or she examines the individual case and not the whole of the results. On this ground the Italian law is silent and the tools lie in Articles 10 and 14 of the European regulation, which, however, will apply to Annex III systems only from December 2027. A similar consideration applies to scoring systems applied to persons, which we addressed when discussing digital cities and the prohibition of social scoring.
Conclusions
The resulting picture is less dramatic and more demanding than it is commonly represented. The AI Act places those systems among high-risk systems, the CEPEJ excludes their binding nature, the Italian law reserves interpretation, application and the assessment of evidence to the judge. The critical point is not the existence of the rule, but its effectiveness, which depends on three conditions that are today absent or incomplete: the training of those who decide, the tools to see what the system has done and on what data, and a timetable that for the high-risk category arrives in December 2027 while use in judicial offices already stands at 46 per cent. Between the adoption of the tools and their regulation there are therefore two years without regulatory coverage.
Author: Valentina Grazia Sapuppo
