Case law invented by AI: what the 2026 French decisions say
From warnings in late 2025 to fines in 2026: the French decisions on AI-hallucinated citations we could verify, what the judges actually wrote, and a verification routine for every brief.
Since December 2025, French courts have started saying it in their decisions: the case law cited before them does not exist, or does not say what it is claimed to say. Here are the decisions we were able to verify, what the judges actually wrote, and a verification routine that takes a few minutes per brief.
1. The decisions
We only cite decisions whose reference and content are confirmed by the full text or by at least two concurring commentaries (in particular Lexbase, 1 April 2026 and Morgan Lewis, 18 March 2026).
| Decision | What the court noted | Outcome |
|---|---|---|
| TJ Périgueux, 18/12/2025, n° 23/00452 | Invites counsel to check that the references found are not "hallucinations". | Warning, no sanction. |
| TA Orléans, 29/12/2025, n° 2506461 | A series of references to decisions that do not exist; the same warning. | No sanction; the claimant succeeded on other grounds. |
| TA Rennes, ord., 28/01/2026, n° 2506364 | A claim "which, moreover, was clearly drafted with a generative AI tool" (our translation; commentary). | Rejected by order (art. R. 222-1 CJA): no argument was precise enough to be assessed. |
| CAA Bordeaux, ord., 26/02/2026, n° 25BX02906 | Four cited decisions concern other matters, two do not exist (full text). | Appeal rejected; warning to counsel. |
| TA Grenoble, ord., 04/06/2026, n° 2410230 | Inapplicable texts and non-existent court decisions in a very long AI-drafted brief. | €200 fine for abusive proceedings (source, source). |
| TA Rennes, ord., 19/08/2026, n° 2508168 | A claim clearly written with AI, without the most basic details, repeating an earlier rejected claim. | €500 fine for abusive proceedings (art. R. 741-12 CJA) (source). |
2. What the judges actually wrote
The Bordeaux administrative court of appeal order is worth reading in full, because it shows the two ways a citation can fail.
“Counsel for Mr B. should be reminded of the need to check the court decisions cited, which were not produced, before referring a matter to the court. [...] It is therefore appropriate to invite the applicant's counsel to check in future that references found by whatever means do not constitute a "hallucination" or a "confabulation".”
Of the six decisions cited in that case, four exist but concern something else entirely (training contracts, abusive requests to the administration, jurisdiction over international relations, freedom of worship), and two do not exist. The invented reference is the best-known error; the real but irrelevant one was, in this case, the more frequent, and it is the harder to catch without opening the decision.
Two observations from the decisions as a whole. First, none of them rejects a claim because it was written with AI: the judges rely on ordinary procedural grounds (lack of precision, manifestly unfounded appeal, abusive proceedings) that the AI-generated drafting merely made visible. Second, the tone is hardening: from warnings in late 2025 to fines in mid-2026.
3. A verification routine
The CNB's 2026 guide calls for "systematic verification of the results produced by AI" (our translation). In practice, for every reference in a brief:
- Open the decision at the source, not a summary: Légifrance, Judilibre for the judicial courts, administrative justice open data, or your usual database.
- Check the identifiers: court, formation, date and number must all match the same decision. A real number with the wrong date is a warning sign.
- Find the passage you rely on, word for word. In the Bordeaux case, most of the bad citations were real decisions on another subject.
- Check it is still good law: later reversal, change in the text applied.
- Produce what you cite. The Bordeaux court noted that the decisions were "not produced". Attaching them forces the check.
Want to see how this works on your own files?