AI Hallucinations in Court: Siracusa Tribunal Sanctions the Lawyer

A lawyer uses a generalist LLM to build their defense brief and cites four Supreme Court rulings complete with quoted holdings. The problem? Those rulings, as cited, don't exist. The Tribunal's verification on the Supreme Court's CED database is immediate, the conclusion unequivocal: AI hallucinations. The sanction is twofold — toward the opposing party and toward the court's fine fund.

01 — The case: a claim lost from the start

The case concerns a landlord company suing the representative of an unrecognized association for damages, invoking Art. 38 of the Civil Code for lost profits (unpaid rent for the remaining period) and loss of goodwill compensation — claims already declared inadmissible on appeal because they were raised for the first time at that stage.

The defendant promptly raises the limitation defense under Art. 1957 of the Civil Code: established and unanimous case law equates the association representative's liability to a statutory guarantee analogous to a surety, with the consequent applicability of the six-month term. The plaintiff had only filed the specific claims in 2018, over two years after expiration.

02 — Gross negligence: failing to supervise AI work and attributing entirely non-existent passages to case law precedents

To counter the limitation defense, the lawyer files a brief under Art. 171-ter, n. 1, of the Code of Civil Procedure, quoting four Supreme Court holdings complete with section, date and registry number: Cass. n. 1216/2000, Cass. n. 8379/2006, Cass. n. 14795/2003, Cass. n. 4553/2004.

"None of the aforementioned rulings contain the quoted passages. The decisions [...] in their authentic formulation, deal with matters extraneous to the issue. In all four cases the quoted passages find no match in any existing ruling."

The Tribunal, in order to identify a logical reason behind these non-existent citations, rules out every alternative hypothesis: database malfunction (databases don't generate text, they index authentic documents), memory error (since these are newly constructed holdings, not wrong numbers), deliberate fabrication (implausible given disciplinary consequences). The only rational conclusion: use of a generalist LLM without verification against primary sources.

It should be noted, however, that even this hypothesis (the most plausible one) still carries disciplinary consequences, not to mention the potential violation of privacy regulations.

03 — The Tribunal's reasoning: a "notorious fact"

The most relevant part of the reasoning — destined to set precedent — is the qualification of the hallucination phenomenon as a notorious fact expected to be known by every legal professional. Large Language Models are not legal databases. They generate statistically plausible text sequences based on training parameters, without ordinary access to verified knowledge bases. "Hallucinations" — formally plausible but substantially false content — are a structural and well-known characteristic of these systems.

Using them without verification against primary sources (CED, professional legal databases, official repositories) constitutes gross negligence. These are not typos or inaccuracies: these are non-existent precedents that force judges and opposing parties into extraordinary verification work.

The Tribunal, in short, does not stigmatize the use of AI per se — indeed, the decision is perfectly in line with the trend of other courts that have excluded automatic sanctions for the mere use of AI tools in drafting documents (Trib. Prato). What is punished is the total absence of professional supervision over the output.

04 — The landscape: not an isolated case

The Siracusa ruling fits into a rapidly evolving case law trend. Italian courts are progressively building a standard of professional diligence in AI use, articulated on a continuum from diligent and faultless conduct (correct use with supervision) to aggravated liability (unverified non-existent citations).

05 — Implications for AI Act and professional liability

The regulatory reading of the ruling is of great interest for AI Act compliance professionals. Regulation (EU) 2024/1689 — fully applicable from August 2026 — does not directly classify generalist LLMs as high-risk systems in the justice domain (Annex III), but introduces general transparency obligations for general purpose AI models (GPAI) and liability for deployers.

Articles 13 (transparency) and 14 (human oversight) of the AI Act require deployers of high-risk AI systems to ensure that output is monitored and verified by qualified personnel. The Siracusa Tribunal's logic — mandatory professional supervision over AI output — anticipates exactly this regulatory standard, making it already enforceable today as a professional deontological obligation, regardless of the AI Act's full application.

On the civil liability front, the ruling reinforces the thesis — already supported by part of legal scholarship — that lawyers who produce court filings with LLM-generated content are liable for quality control according to ordinary professional diligence standards. The reference norm is not the AI Act, but forensic ethics and Art. 96 CPC: the European regulation merely crystallizes a standard already emerging from case law.

06 — The verification protocol: what to do (concretely)

The Siracusa Tribunal does not ban AI use. It identifies an operational procedure that every professional should adopt:

This article is for informational purposes and does not constitute legal advice. The positions expressed reflect Tomato Blue RegTech's analysis. For specific assessments, consult a qualified professional. © 2026 Tomato Blue.