EU judicial AI decision making

Italy's Cassation Court Has Parked the Judicial-AI Question, and the Delay Is a Chance to Get the Rules Right

A pending Italian referral on judges' AI use shows why disclosure and human accountability, not blanket bans, should anchor EU judicial-AI rules.

Judicial AI in Italy and the EU People of Internet Research · EU 2 Dec 2027 Annex III obligations apply Delayed high-risk date under the D… 2 Appeals joined by the Court The Court joined the appeals and r… Jul–Aug 2026 Cassazione Copilot trial length Tax-section screening of appeals, … peopleofinternet.com
Judicial AI in Italy and the EU People of Internet Research · EU 2 Dec 2027 Annex III obligations apply 2 Appeals joined by the Court Jul–Aug 2026 Cassazione Copilot trial le… peopleofinternet.com

Key Takeaways

On 28 October 2026 Italy's Corte di Cassazione published interlocutory order no. 27957/2026, from its First Civil Section, chaired by Mauro Di Marzio. According to Il Sole 24 Ore's legal commentary, a party asked the Court to refer to the Sezioni Unite the "risk of using artificial intelligence" in supporting or drafting the analyses, notes and drafts behind judgments. The Court did not answer. It joined two related appeals, in a dispute involving the Città Metropolitana di Catania, and reserved "any assessment" until after the joinder.

That restraint matters. Studio Cataldi reports that the Court did not establish any actual use of automated systems by the lower-court judges, and took no position on lawfulness. So this is not a ruling about AI-drafted judgments. It is a pending question in a member state's highest civil court, which makes it a useful test of how Europe should think about AI in courts.

The strongest case for tight control

The case for strict rules is serious. A judgment is the state's most coercive speech act, and litigants are entitled to a reasoned decision from a human judge. Language models can fabricate authorities and can embed bias invisibly, and a draft that looks fluent is easy to rubber-stamp. If a judge's reasoning was shaped by a tool no one disclosed, the party cannot challenge it. The EU legislature reached a similar view. Regulation (EU) 2024/1689, the AI Act, treats systems used by or on behalf of a judicial authority to help research and interpret facts and law, and apply the law, as high-risk under Annex III. Recital 61 explains that the classification addresses risks of bias, errors and opacity, while excluding purely ancillary administrative activities.

Why a blanket answer would be a mistake

The risk the appellants point to is real. But the response should track what the tool does and who answers for the result, not whether software touched the draft.

AI-assisted work in a court runs along a spectrum. At one end, a model summarises a 200-page file, retrieves case law or checks citations. At the other, it proposes a disposition. The first uses resemble the research clerks, bench memos and databases courts already use. The second raises the legitimacy concerns the AI Act targets. Treating both as a defect that voids a judgment would push use underground, which is the worst result for transparency.

Italy's own judiciary is already experimenting. Il Post reported in September 2026 that between July and August the Cassazione trialled Microsoft Copilot in its tax section to screen appeals. The output was a summary card checked by an operator and a magistrate, and the tool did not evaluate precedents. That is a bounded, supervised design, and it is the kind of use courts need to be able to adopt. Backlogs are a justice problem too, since a decision that arrives years late is itself a denial of rights.

There is also an evidentiary problem. As the Il Sole commentary notes, one of the contested judgments dates from 2021, before generative AI was widely available, so an objection that it was machine-drafted is implausible for that decision. An appellant who says "AI may have been used" and nothing more is asking a court to infer a vice from a possibility. The Court's decision to join the appeals first and reserve the question is a sensible refusal to build doctrine on speculation.

The regulatory clock

The compliance timeline is moving too. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was adopted on 8 July 2026 and published in the Official Journal on 24 July 2026. Under it, the Annex III high-risk obligations apply from 2 December 2027, as the Il Sole commentary also notes. That is more than a year of runway before the high-risk judicial-AI regime binds anyone.

The delay is not an argument for doing nothing. Courts are not waiting for the AI Act, and neither are litigants. The gap invites a patchwork, with one chamber tolerating tools, another treating them as grounds for nullity, and no shared standard on disclosure.

What proportionate rules would look like

If the Sezioni Unite take the question, or if the legislature acts first, four principles would serve both justice and innovation.

This approach keeps the legitimate core of the appellants' worry, the right to a reasoned human decision, without freezing tools that can shorten cases and reduce errors.

What to watch

Two things are worth watching. The first is whether the First President sends the question to the Sezioni Unite, since referral is not automatic. The second is whether the Court, if it rules, says anything about disclosure or only about nullity. A nullity-only rule would reward tactical allegations and punish transparency. A disclosure-and-accountability rule would line up with what the AI Act is trying to do from December 2027.

Europe's high-risk classification for judicial AI is a reasonable starting point. The risk now is that national courts improvise before the harmonised rules arrive. The best outcome in Rome would be a measured, principle-based answer that treats judges as accountable professionals who may use tools, not as suspects whose every draft must be audited for software.

Sources & Citations

  1. Regulation (EU) 2024/1689 (AI Act), EUR-Lex
  2. Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex
  3. Il Sole 24 Ore NT+ Diritto on order 27957/2026
  4. Studio Cataldi on the Cassazione order
  5. Il Post on Cassazione's Copilot trial