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.
On this page
AI hallucinations happen because a language model predicts plausible text rather than retrieving records: to the model, a citation that looks right is as good as one that is right. The result is measurable. By 14 September 2026 Damien Charlotin's database listed 2,041 court decisions worldwide involving hallucinated material, 811 of them involving lawyers, and the rate has settled at roughly eight new decisions a day.
The mechanism: prediction, not retrieval
A model such as Claude Opus 5 or GPT-6 Astra generates a reply one token at a time, each token chosen as the most probable continuation of everything so far. Ask for "the leading Second Circuit authority on equitable tolling in FCRA cases" and the model has seen thousands of sentences of that shape. It knows what such an answer looks like: a party name, a reporter citation, a year, a pinpoint, a holding phrased in the right register. It does not know whether any particular combination exists, because existence is not something it ever stored. (The mechanics are set out in how large language models work.)
Two things make this worse in legal work. Citations are strings of unusual tokens with weak statistical ties to their content, so the model has learned the shape of "[2025] EWHC 1383 (Admin)" far better than the substance. And the training process rewards answering. OpenAI's researchers wrote in September 2025 that models hallucinate because training and benchmarks reward guessing over abstaining, and that rewarding "I don't know" reduces the rate substantially (OpenAI, 5 September 2025). Models are also sycophantic: Stanford's 2026 AI Index reports that when a false premise is framed as the user's own belief, accuracy on one benchmark fell from 98.2% to 64.4% for GPT-4o (Stanford HAI). If you tell the model what you hope the law says, it will help you find it.
The numbers: what 2,041 decisions look like
Charlotin's inclusion bar is strict: a court must find, or clearly imply, that a party relied on hallucinated material. Allegations do not count, so the author calls the figure "necessarily an undercount" (AI Hallucination Cases database).
| Dimension | Count (14 September 2026) |
|---|---|
| Total decisions | 2,041 |
| United States | 1,395 |
| Canada | 217 |
| Australia | 111 |
| United Kingdom | 69 |
| Israel / Brazil / India / Italy / France / Germany | 57 / 41 / 15 / 15 / 14 / 11 |
| Pro se litigants | 1,174 |
| Lawyers | 811 |
| Judges | 32 |
| Experts | 15 |
| Fabricated case law (instances) | 1,690 |
| Misrepresented citations | 854 |
| False quotations | 548 |
The growth curve tells its own story: about 87 decisions in May 2025, 486 by October 2025, 712 by 29 December 2025 (90% of them that year), 1,227 in early April 2026, 1,598 on 9 June 2026 and 2,041 in mid-September. Scientific American reported in May 2026 that the exponential phase has given way to a plateau of 350–400 decisions a quarter and attributed the persistence to "cognitive surrender", the tendency to reverse a correct judgment when a machine disagrees (Scientific American).
Two readings of the table are both correct. Most entries are litigants in person, so the headline overstates the lawyer problem. And the lawyer entries are the ones that end in fines, suspensions and referrals: $110,204.38 in Couvrette v. Wisnovsky (D. Or., 12 December 2025), $15,000 per attorney plus fees and double costs in Whiting v. City of Athens (6th Cir., 13 March 2026), an interim suspension in Nebraska ordered on 15 April 2026 for a brief with 57 of 63 citations defective, and a striking-off by the Solicitors Disciplinary Tribunal in SRA v Kumar (decided 25 August 2026). The full run of cases is in AI hallucination sanctions: the cases every lawyer should know.
Three kinds of hallucination, and which one you will miss
Charlotin's categories map onto three failure modes, in ascending order of difficulty to catch.
1. The fabricated case. The authority does not exist. This is the Mata v. Avianca (S.D.N.Y. 2023) pattern, six invented cases and a $5,000 sanction, and it is the easiest to catch: search the citation and nothing comes back. It still accounts for 1,690 instances because people do not search.
2. The false quotation. The case exists; the quoted words do not appear in it. Fletcher v. Experian (5th Cir., 18 February 2026) involved 16 fabricated quotations and a $2,500 fine; the tools reportedly used were vLex and CoCounsel, not a consumer chatbot. Catching this requires opening the judgment and finding the passage, which is why it survives a surface check.
3. The misrepresented authority. The case is real, the quotation may even be real, and the case does not stand for the proposition. This is the 854-instance category and the one Clearbrief's founder has said is often "not easy to spot with the human eye". HAQQ's June 2026 benchmark found that 24% of 3,000 answers across every model "cited or misapplied law that didn't support claims" (HAQQ). Catching it means reading the case, not just finding it.
Ayinde v Haringey [2025] EWHC 1383 (Admin) contained all three: five authorities that did not exist, plus quotations and propositions attributed to real ones. The Divisional Court's response, examined in Ayinde v Haringey explained, was to hold that "freely available generative AI tools" are "not capable of conducting reliable legal research" and to place responsibility on heads of chambers and managing partners.
What the benchmarks actually show
Marketing copy says "hallucination-free" or "grounded in authoritative sources". The independent and semi-independent measurements say something more specific.
| Study | Date | What was tested | Hallucination finding |
|---|---|---|---|
| Stanford RegLab, "Hallucination-Free?" | May 2024 (peer-reviewed 2025) | 202 expert-scored queries on commercial tools | Lexis+ AI 17%, Westlaw AI-Assisted Research 33%, GPT-4 43%; fully accurate answers 65% / 41% / 19% |
| Vals Legal Research study | 23 October 2025 | 210 US research questions | Counsel Stack 81%, Alexi 80%, ChatGPT with search 80% vs lawyer baseline 71%; all tools lost 11–14 points on multi-jurisdictional questions; Westlaw, Lexis and vLex declined to take part |
| HAQQ benchmark | 5 June 2026 | 3,000 graded answers, 300 tasks, 51 practice areas | Citation hallucination GPT-5.5 3%, Claude Opus 4.8 about 12%, Grok 4.3 about 14%, o3 32%, Mistral Large 64%; 24% of all answers misapplied law |
| UPenn deep-research study | 3 April 2026 | Search-augmented models and deep-research agents | 3–13% of cited URLs fabricated; agents hallucinate more because they cite more |
| Vectara summarisation leaderboard | 11 May 2026 | Summarising a supplied document | Best models 1.8–3.3% |
Sources: Stanford RegLab; LawSites on Vals; HAQQ; UPenn, arXiv; Vectara.
Three lessons. Rates have fallen sharply on structured tasks (summarising a document you supply is now very reliable) and remain material on open research. The best model on one axis is not the best on another: in HAQQ's test GPT-5.5 had the lowest hallucination rate but was the slowest (134 seconds) and dearest ($0.082 a task), while Claude Opus 4.8 won the overall score. And the vendors with the most to lose have declined public tests since 2024, which is why Stanford's figures, old as they are, remain the last independent measurement of Lexis and Westlaw's tools. How to read all of these is covered in legal AI benchmarks explained and the model-by-model answer in which AI has the lowest hallucination rate for law.
Why retrieval helps and does not cure
Retrieval-augmented generation gives the model real documents to read before it answers. It is the reason Westlaw-grounded CoCounsel, Shepard's-grounded Protégé and citation-checking tools exist, and it is the reason legal tools do better than raw chatbots. It fails in three predictable ways: the retriever finds the wrong passage; the model over-extends a right passage to a proposition it does not support; or the model ignores retrieval and answers from memory. Stanford's authors found vendor claims of "hallucination-free" RAG overstated in 2024, and nothing published since has shown a zero rate.
Cork v Smith [2026] EWHC 1199 (Ch) (4 June 2026) is the cautionary example: a junior associate at Pinsent Masons asked an AI to research insolvency law, it invented "Insolvency Rule 12.37(5)", the tool itself flagged that the answer might be inaccurate, and the filing went in anyway. The judge described a "cavalier attitude"; the firm was referred to the SRA (UK Human Rights Blog). Wadsworth v. Walmart (D. Wyo., 24 February 2025), where 8 of 9 cited cases were fake, involved Morgan & Morgan's internal AI platform. Grounding lowers the odds. It does not change who signs the brief.
A verification routine that survives a bad day
The routine below is short because it has to be done under deadline pressure, which is when the cases in the database happened.
| Check | Question | Where | Pass condition |
|---|---|---|---|
| Existence | Does this authority exist with this citation? | Westlaw, Lexis, BAILII, CourtListener, RIS (Austria), juris (Germany) | Found at the citation given, correct parties and year |
| Support | Does it say what the memo says, at the pinpoint given? | The judgment itself | You have read the paragraph and it supports the proposition without stretching |
| Treatment | Is it still good law? | KeyCite, Shepard's, the court's later citations | No adverse treatment on the point relied on |
| Provenance | Did the AI supply this, or did you? | Your research log | Every AI-supplied authority is marked and has passed the three checks |
Build the check into the prompt so there is less to catch afterwards:
You are assisting a litigation associate. Work ONLY from the authorities pasted below;
do not cite, name or allude to any case, statute or rule that is not in the pasted material.
If the pasted authorities do not support a proposition, say "NO SUPPORT IN PROVIDED SOURCES"
rather than filling the gap. For every proposition, give the authority, the pinpoint
paragraph number and a verbatim quotation of no more than 25 words. At the end, list any
propositions you could not support.
[PASTE VERIFIED AUTHORITIES HERE]
Three rules complete the routine. Do not ask the model whether its cases are real; the lawyers in Mata did that, and a New Jersey lawyer was sanctioned a second time in April 2026 for "using AI to verify AI". Log what the AI contributed, because courts have consistently punished the cover-up more than the error (the Fifth Circuit told counsel in Fletcher that owning up early would have meant a lighter penalty). And apply the routine to your opponent's authorities as well: in Noland v. Land of the Free (Cal. Ct. App., 12 September 2025) the respondent's counsel was criticised for failing to notice the appellant's fabricated citations. The step-by-step protocol, with a research-log template, is in how to verify AI legal research; the safe way to use AI at the start of a research task is in AI legal research in 2026.
What to do before your next filing
Adopt the three-check table as a firm rule and put it in the AI policy. Require every AI-assisted brief to carry a research log naming which authorities the model supplied. Use the prompt above, or a firm version of it, for every research task, and switch to a Westlaw- or Lexis-grounded tool for anything that will be filed. Run a citation checker (Clearbrief's Cite Check Report, BriefCatch, Westlaw Quick Check) as a second pass, never as the first. Module 2 of the AI Lab for Lawyers course runs this verification routine live on real research prompts, including the moment the model invents a case, so that the failure is something you have seen rather than something you have read about. The glossary entry on hallucination is the one-page version to hand to trainees.
Frequently asked questions
Why does AI make up case citations?
A language model writes the most probable next words, and after a sentence like 'the leading authority is' the most probable continuation is a plausible case name, citation and holding. Nothing in the model checks that the case exists. OpenAI's own researchers add that training rewards guessing over saying 'I don't know', so confident invention is the learned default.
How often do legal AI tools hallucinate?
It depends on the tool, the task and the year. Stanford's RegLab measured 17% for Lexis+ AI and 33% for Westlaw AI-Assisted Research in 2024. HAQQ's June 2026 benchmark found citation hallucination rates from 3% (GPT-5.5) to 64% (Mistral Large), and 24% of all 3,000 answers misapplied real law. No tool tested has reached zero.
Does retrieval-augmented generation stop hallucinations?
It reduces them by giving the model real sources to work from, but the model can still misread, over-extend or ignore what it retrieved. Fletcher v. Experian (5th Cir. 2026) involved fabricated quotations from paid legal tools, and a Pinsent Masons associate in Cork v Smith filed an invented insolvency rule that the AI itself had flagged as possibly inaccurate.
What is the safest way to check AI-generated legal research?
Open every authority in Westlaw, Lexis, BAILII or the court's own database and confirm three things: it exists, it says what the AI claims at the pinpoint given, and it has not been overruled or distinguished. Do this yourself. Courts have sanctioned lawyers who asked an AI to verify another AI's citations, and the duty to check is described as nondelegable.
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
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
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
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.
5 min read
Questions
Which AI Has the Lowest Hallucination Rate for Legal Work?
The lowest hallucination rate AI for legal work as of September 2026: GPT-5.5 at 3% on HAQQ, Claude Opus 4.8 best overall, and why no model reaches the zero courts demand.
6 min read
Use Cases
AI Legal Research in 2026: What Works, What Hallucinates
AI legal research now beats a lawyer baseline on accuracy yet misapplies law in 24% of answers. Which tools are grounded, where errors hide, a routine that holds.
11 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