AI hallucinations: what the first French court decisions teach lawyers

A hallucination is a false answer, presented with the same confidence as a correct one. In law, it often takes the form of a court decision that does not exist, a wrong case number or a text cited outside its scope. The risk comes from its plausibility: the reference looks like a genuine reference.

The phenomenon is not limited to consumer tools. A study by Stanford researchers, published as a preprint in May 2024, measured hallucination rates of 17% to 33% for two legal research tools from major publishers, while placing them below the level of a general-purpose model such as GPT-4. These figures are old and the products have changed. They do not describe current tools, but they show that a document database reduces the risk without eliminating it.

The known cases are concentrated over a few months. We present them based on published commentary, in particular that of the law firm Morgan Lewis in March 2026:

• Administrative Court (tribunal administratif, TA) of Grenoble, 3 and 9 December 2025 (No. 2509827 and No. 2512468). The judge notes the lack of clarity of filings likely drafted with generative AI, without this affecting the outcome of the decision.
• Judicial Court (tribunal judiciaire, TJ) of Périgueux, 18 December 2025 (No. 23/00452). The court asks the party and its counsel to check in the future that their references are not hallucinations.
• TA Orléans, 29 December 2025 (No. 2506461). The court notes some fifteen false references in the filings of a claimant represented by a lawyer. It asks counsel to check their sources, without a penalty.
• TA Rennes, 28 and 30 January 2026 (No. 2506364 and No. 2600610). Filings clearly drafted with AI are dismissed, for lack of detail in one case and before a court without jurisdiction in the other.
• Administrative Court of Appeal (cour administrative d’appel, CAA) of Bordeaux, 26 February 2026 (No. 25BX02906). The interim relief judge reiterates the need to check the decisions cited before bringing a case. The judge did not dismiss the application on that ground, according to Le Monde du Droit.

The most widely quoted wording comes from the administrative courts: counsel is asked to check that the references are not a “hallucination” or a “confabulation.” It targets the reference however it was found. The focus is therefore on review, not on a tool or a provider.

Yes, but so far they target the abusive nature of a proceeding. They rely on Article R. 741-12 of the French Code of Administrative Justice, which allows a fine of up to 10,000 euros:

• TA Grenoble, 4 June 2026 (No. 2410230). A 200-euro fine against a claimant, for an application targeting decisions that had become final, with a brief of more than 300 pages citing inapplicable texts and nonexistent decisions.
• TA Rennes, 19 August 2026 (No. 2508168). A 500-euro fine for an application repeating a claim that had already been dismissed. The available commentary does not all present the role of AI in the same way.
• TA Nîmes, 11 September 2026 (No. 2604053). A 1,000-euro fine for an application prepared with AI, very long and with no justified urgency, according to Que Choisir.

According to the commentary, the use of AI is one element of the case file in these matters, not a separate offense. The Grenoble judgment concerns a claimant not represented by a lawyer. An article dated August 20, 2026 stated that no disciplinary penalty had been imposed in France on a lawyer for a hallucination at that time. This finding should be checked again as of the date you read this.

The database maintained by researcher Damien Charlotin listed 2,266 decisions worldwide as of November 5, 2026, including 14 for France. This total should be read with caution. The database depends on the decisions identified and published, and its author presents it as a work in progress.

In January 2026, Next still counted only three French cases in the same database. The increase may reflect more use as well as better detection. The available figures do not make it possible to decide between the two.

These decisions concern litigation filings, not contracts. None of those listed here involves a contract or a legal department. The comparison therefore remains a line of reasoning, not a rule.

The ethics rules cited apply first to attorneys. Morgan Lewis points to the principles of competence, diligence and prudence in Article 1.3 of the National Internal Rules of the French legal profession (Règlement intérieur national), and to the March 2026 guide of the French National Bar Council (Conseil national des barreaux), which mentions a disciplinary risk for unverified content. An in-house lawyer does not fall under this disciplinary framework, but the same logic of verification applies to their work.

In a company, the risk simply takes another form. It may be a memo to management based on a nonexistent ruling, a summary that attributes a missing clause to the contract, or a reply to a partner based on an amended text. An error of this kind can influence a decision without ever coming before a judge.

These habits apply to any tool. They require few resources, and their cost remains low compared with the error avoided:

• Open each reference. A cited decision is checked on Judilibre, the open data of the administrative courts or Légifrance. If it cannot be found, it is not used.
• Read the source, not just confirm it exists. A real decision can be cited for a holding it does not adopt.
• Check the date and the version. An article may exist and have been amended since.
• Go back to the contract. Any statement about a clause is checked in the corresponding passage, in the right version.
• Name the reviewer. A named person takes responsibility for checking important documents.
• Train business teams. An AI answer passed on as is to a partner has not been approved by the legal department.

These habits do not require giving up AI. The Paris Bar and the French National Bar Council, cited by Morgan Lewis, frame the issue in terms of verification and responsibility.

An AI can make mistakes, including in Docform: the decision stays with the teams. The assistant cites the source passage for each point it flags, to make it easier to go back to the contract, and it queries Légifrance and Judilibre for legal research. The lawyer checks, corrects and approves. How the assistant works is detailed on the dedicated page.