Glossary
Hallucination (AI) — Definition for Lawyers
An AI hallucination is fluent, confident, false output: an invented case, a misread holding or a fake quote. 2,041 court decisions now involve one. What to know.
On this page
A hallucination is output from an AI model that is fluent, confident and false: a case that does not exist, a real case cited for a proposition it does not support, or a quotation nobody wrote. Courts have now recorded 2,041 decisions involving hallucinated material worldwide, 811 of them attributable to lawyers, according to Damien Charlotin's database as of 14 September 2026.
The three kinds a court will notice
Charlotin, a research fellow at HEC Paris, only logs a case when a court finds or clearly implies that hallucinated material was relied on, so his AI Hallucination Cases database is an undercount. It still went from about 200 entries in mid-2025 to 712 on 29 December 2025 and 2,041 on 14 September 2026, adding roughly eight decisions a day in June 2026. The United States accounts for 1,395, Canada 217, Australia 111, the UK 69, Germany 11. Pro se litigants are behind 1,174 entries, lawyers 811, judges 32 and expert witnesses 15.
The database splits hallucinations into three types, and the split matters more than the total.
| Type | What it looks like | Instances (14 Sep 2026) | Example |
|---|---|---|---|
| Fabricated authority | A case, statute, rule or article that does not exist | 1,690 | Cork v Smith [2026] EWHC 1199 (Ch): an AI invented "Insolvency Rule 12.37(5)", the associate filed it, and the firm was referred to the SRA |
| Misrepresented authority | A real case cited for something it does not say | 854 | On the HAQQ benchmark (June 2026), 24% of 3,000 answers across every model cited or misapplied law that did not support the claim |
| False quotation | Words attributed to a judgment that do not appear in it | 548 | Fletcher v. Experian (5th Cir. 2026): 16 fabricated quotations, traced to vLex and CoCounsel, $2,500 fine |
The first type is the easiest to catch, because a search for the citation returns nothing. The second passes an existence check and fails only when someone reads the judgment. Europe is not exempt: on 7 October 2025 the Austrian OGH dismissed an AI-drafted Nichtigkeitsbeschwerde (14 Os 95/25i) containing invented OGH decisions without reaching the merits. The full list of consequences is in the sanctions case guide.
Why it persists, even in legal-grade tools
OpenAI's own researchers explained the mechanism in September 2025: training and benchmarks reward guessing over "I don't know", so a model that has never seen the case you need produces the case that is statistically most likely to exist. Stanford's 2026 AI Index measured hallucination rates from 22% to 94% across 26 leading models on a factuality benchmark, and found that when a false premise is framed as the user's own belief, accuracy collapses, GPT-4o from 98.2% to 64.4%. Tell a model what you think the case holds and it will tend to agree with you.
Grounding the model in a database helps but does not cure. Stanford RegLab's "Hallucination-Free?" study tested the research tools directly and found Lexis+ AI hallucinating 17% of the time and Westlaw AI-Assisted Research 33% on 2024 versions, against 43% for GPT-4. Lacey v. State Farm (C.D. Cal. 2025) cost Ellis George and K&L Gates $31,100 for a brief built with CoCounsel, Westlaw Precision and Gemini. Adding web search does not close the gap either: a University of Pennsylvania study (April 2026) found search-augmented deep-research agents fabricating 3% to 13% of the URLs they cited, and hallucinating more because they cite more. The mechanics of grounding, and where it fails, are set out under retrieval-augmented generation; the model-by-model figures are in which AI has the lowest hallucination rate.
The pattern in the sanctions cases is not a lack of tools. In Cork v Smith the AI flagged its own answer as possibly inaccurate and the associate filed anyway. In Mata v. Avianca (S.D.N.Y. 2023) the lawyers asked ChatGPT whether its cases were real, and in April 2026 a New Jersey lawyer was sanctioned a second time for "using AI to verify AI". Charlotin's tracker treats asking a model to validate itself as an aggravating factor.
What a lawyer can do about it
The Divisional Court in Ayinde v Haringey [2025] EWHC 1383 (Admin) held that freely available generative AI tools "are not capable of conducting reliable legal research" and told heads of chambers and managing partners that the responsibility is theirs. The working protocol that follows from the case law has three checks per citation, and none of them is delegable to the model:
- Existence. Open the authority in Westlaw, Lexis, BAILII, RIS or the court's own database. Not in a chatbot.
- Support. Read the paragraph relied on and confirm it says what the draft says it says. This is the check that catches the 854 misrepresented authorities.
- Treatment. Run KeyCite, Shepard's or the local equivalent to confirm the case is still good law.
Two further rules cost nothing. Never paste the model's output back into a model and ask whether it is correct. And if a fake citation reaches a court despite everything, say so at once: Norton Rose Fulbright's June 2026 review of the case law notes that candour mitigates and evasiveness aggravates, which is the difference between no sanction in Gamez v. County of Fresno (E.D. Cal. 2026) and $15,000 per lawyer in Whiting v. City of Athens (6th Cir. 2026) after counsel refused to answer the court's questions.
The step-by-step version is in the citation-checking protocol, and the fuller treatment of causes and controls in the guide to AI hallucinations in legal practice. Module 2 of the AI Lab for Lawyers course covers legal research and drafting, which is where this protocol belongs.
Frequently asked questions
What is an AI hallucination in legal terms?
It is any output a model presents as fact that is not true: a case, statute or rule that does not exist; a real authority cited for a proposition it does not support; or a quotation that does not appear in the source. The word is a term of art from machine learning, not a diagnosis. Courts treat the resulting filing exactly as they would a human fabrication, under existing rules such as FRCP 11 and the duty of candour.
Do legal AI tools like Westlaw and Lexis still hallucinate?
Less than general chatbots, but yes. Stanford's RegLab study measured 17% for Lexis+ AI and 33% for Westlaw AI-Assisted Research on 2024 versions and found vendor claims of hallucination-free retrieval overstated. In 2025 and 2026, lawyers were sanctioned for filings produced with CoCounsel, Westlaw Precision and vLex. Grounding shifts the failure from invented cases to real cases misread, which is harder to spot.
Which AI has the lowest hallucination rate for legal work?
On the vendor-run HAQQ benchmark of June 2026, GPT-5.5 had the lowest citation hallucination rate at 3%, but it was also the slowest and most expensive model tested, and 24% of all 3,000 answers across every model cited or misapplied law that did not support the claim. No model tested reached zero, and a court expects zero. Treat the benchmark as a ranking, not a licence to skip verification.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Guides
AI Hallucinations in Legal Practice: Why They Happen and How to Catch Them
AI hallucinations in legal practice: why models invent cases, what 2,041 court decisions and the HAQQ benchmark show, and a verification routine that holds.
9 min read
Playbooks
How to Verify AI Legal Research: A Citation-Checking Protocol
Verify AI legal research with a five-check protocol: open every authority, match the quote, confirm holding and status, log it. Template, time budgets, tools.
11 min read
Regulation & Ethics
AI Hallucination Sanctions: The Cases Every Lawyer Should Know (2023–2026)
AI hallucination cases reached 2,041 in Charlotin's database by September 2026, 811 involving lawyers. Every sanction from Mata to Couvrette's $110,204, with lessons.
12 min read
Regulation & Ethics
Ayinde v Haringey Explained: The UK's Leading AI Hallucination Judgment
Ayinde v Haringey [2025] EWHC 1383 (Admin) explained: the fake cases, the Divisional Court's sanctions ladder, the duty on heads of chambers and how the SRA built on it.
11 min read
Questions
Can AI Do Legal Research Without Hallucinating? The Honest Answer
No AI legal research tool is hallucination-free as of September 2026. Measured rates, why grounding does not cure misreading, and the checks that protect you.
6 min read
Glossary
Retrieval-Augmented Generation (RAG) — Definition for Lawyers
Retrieval-augmented generation (RAG) searches a document set first and drafts from what it finds. Every grounded legal tool uses it; here is why errors still get through.
4 min read