Fake AI Citations: A Verification Protocol for Lawyers

In June 2023 a New York lawyer told a federal judge he hadn't realised ChatGPT could invent court decisions. The six cases in his brief looked right: plausible names, docket numbers, even quotations. None of them existed. He, a colleague and their firm were fined $5,000, and Mata v. Avianca became the case everyone cites on this subject.

You would think that settled the matter. It didn't. On 5 November 2026 the public database kept by researcher Damien Charlotin listed 2,266 court decisions in which a court found, or inferred, that a filing relied on hallucinated content, and in 898 of them the party was a lawyer. In early July 2025 the same database stood at more than 160.

This guide is for lawyers who use AI (many of us already do) and who would like to stay out of that database. It rests on what courts and bars in eight countries have actually said, not on vendor checklists. The six-step check below takes minutes per authority. Skipping it has cost lawyers fines, referrals to their bar, a wasted-costs order of AED 282,508 and, in one case reported in Turkey, investigations by both the prosecutor and the bar.

In short

• The lawyer who signs owns every authority in the document, whichever tool drafted it. Courts in the UK, the UAE and Spain have said so in plain words.
• Most fake authorities are simply invented. In the 898 decisions that involve a lawyer, a source was invented in 8 in 10, a real source was misdescribed in almost half, and a quotation was missing from its source in a third. One decision can carry several of these labels.
• Legal-specific AI tools are better than general chatbots, not infallible. In a 2024 Stanford study of more than 200 queries built to probe weaknesses, Lexis+ AI, Ask Practical Law AI and Westlaw's AI-Assisted Research each hallucinated between 17% and 33% of the time. The products have changed since; the principle hasn't.
• Speed and candour help, but they guarantee nothing. A Spanish lawyer who apologised at once avoided a fine. In England a solicitor who admitted the error and reported himself was referred to his regulator all the same, and lawyers who called five fake cases "cosmetic errors" were told they should have been shocked.
• Six checks do the work: does it exist, does it match, is the quotation real, does it support the point, is it still good law, and did you keep a record.

Five ways a citation goes wrong

"Fake case" is the headline, but it is only one of five failures, and the others are harder to spot. The database labels each decision by the kind of error the court found, and one decision can carry several labels.

The last row is the rarest in the database and, in our reading, the most under-counted. A judge notices a case that does not exist long before noticing one that was overturned last spring.

Why does a language model produce all this? It writes the most plausible next words, and a citation is a pattern with very little slack: party names, court, year, number. If the tool has no access to a legal database, or has one and fails to retrieve from it, it completes the pattern. In the German family-law case below, the court found an author attributed to the wrong commentary, a margin number that does not exist and a monograph it could not trace. A German practitioner who read the order suspected a tool without database access and a weak prompt. Tools that do retrieve from a database make fewer errors, but the error changes shape: a real source, wrongly described. And never ask the same chatbot whether its own citations are real. The New York lawyer did, and it said yes.

The six-step check

Run all six on every authority in anything you file or send a client as advice. For internal brainstorming, steps 1 and 2 are the minimum before you rely on anything. The citation discipline we describe in our piece on winning legal arguments is the same idea seen from the drafting side.

1. Exists

Find the authority by its identifier (court, date, number), not by its title, in the court's own database or a commercial one. A title search can return a different case with a similar name and tempt you to stop. Do not use the AI to confirm itself.

2. Matches

Check that court, date, number, parties and the publication reference all agree with what you cited. This is the step that would have helped the Ankara lawyer described below, whose Yargıtay references, according to press reports, could not be found as cited.

3. Quote

Every quotation must appear word for word in the source. Open the document and search for the phrase. 298 of the 898 decisions that involve a lawyer include false quotations (602 across all parties), and a quotation that sounds like a judge is exactly what a model is good at producing.

4. Supports

Read the passage, not only the summary or headnote. Does the court decide the point you say it decides, on facts like yours, as part of its reasoning and not in passing? A real source that was misrepresented is the second most common failure in the database, after outright invention (419 of the 898 lawyer decisions).

5. Status

Is it still law? Look for later appeals, annulments and overruling, and for the version of a statute in force on the date that matters. Common-law systems have citators. In civil-law systems use the consolidated texts and case-law databases of your own jurisdiction (see the list further down).

6. Record

Write one line per authority: who checked it, where, when. It costs almost nothing, and it is your best evidence if a judge or a bar asks. The Spanish bar's 2026 circular makes the same point about traceability.

What courts have actually done

Every row comes from a court's own press office, a legal-press report or a law firm's published analysis, all linked at the end. Amounts are the ones the sources report.

Three patterns run through these cases. First, outcomes differ by court, and good behaviour is no guarantee. The Divisional Court says that when a lawyer puts false citations before a court, with or without AI, a referral to the regulator is likely to be appropriate, and it referred a solicitor who had admitted the error and reported himself. In Spain, art. 247.4 of the civil procedure law (LEC) obliges the court to tell the bar when the conduct may be a lawyer's. In France and Germany the decisions we found only warned the lawyers. Second, where a court has discretion, speed and candour count: Navarra closed its sanction proceedings because the lawyer apologised in writing, twice, straight away. Third, leaders are in the frame. The UK Divisional Court said it will ask in future whether heads of chambers and managing partners made sure everyone understands their duties when using AI, and that, save in exceptional circumstances, "admonishment alone is unlikely to be a sufficient response".

The numbers behind the cases

The database tracks decisions in which a court or tribunal dealt with hallucinated content; in most of them the court only inferred that AI was used (the database marks the tool as "Implied"). It cannot count every fake citation, so the real number is higher, and part of the growth since 2025 is better tracking. The direction is still unmistakable: more than 160 decisions in July 2025, 2,266 in November 2026.

Poland and Argentina follow with 10 entries each, then Spain with 8. The UAE has 2 and Qatar 1; Turkey and Japan have none. The database lists more than 40 jurisdictions in all, and some decisions appear twice (Germany's 11 entries are 8 decisions). Before you read this as an American problem, note two things: American courts publish far more orders and are watched more closely, and the list has no Turkish entry even though an investigation in Turkey was reported in June 2026.

Who relied on the tool? Self-represented litigants in 1,305 decisions, lawyers in 898, judges in 33 and experts in 17. The last two figures matter in practice: run the same six steps on the other side's authorities, on an expert's report, and on any decision you intend to lean on.

A one-page AI policy for your firm

Bars are converging on the same short list, and a small firm can write it in an afternoon.

1. The signature owns it. Name the lawyer who answers for every authority in a document, whoever or whatever drafted it. Courts already do.
2. Approved tools only. List the tools allowed for client work. Consumer accounts and browser extensions stay out. The tests are where data is processed, whether it is used to train models and what the contract says. We go through them in our article on confidentiality and client data.
3. Names stay out of prompts unless the tool is cleared for them. Strip identifiers by default.
4. Verify before it leaves the building. The six-step check for every authority in anything filed or sent as advice, and a second person for filings above a risk line you define: deadline, amount, appellate level.
5. Say it where the rule requires it. Some courts ask for a declaration of AI use (the Federal Court of Canada does, and a decision of November 2026 criticised a litigant for not making one); many do not. Keep a one-page table of the rules of the courts you appear in, and tell clients how you use AI in the engagement letter.
6. Train and supervise. Trainees, paralegals and clients all use these tools. A thirty-minute session and a standing question in file review (which tool, which prompt, which sources?) cost next to nothing.
7. Keep a trace. Log tool, date and verification for each filing.

What your own bar has published, as far as we found it: the ABA's Formal Opinion 512 (29 July 2024, competence, confidentiality, candor, supervision, fees); in France the national bar council's (CNB) practical guide (2024) and ethics guide on AI (March 2026); in Germany the Federal Bar's (BRAK) guidance of December 2024; in Spain the bar council's (CGAE) Circular 3/2026 (April 2026); in Brazil the bar's (OAB) recommendations of 11 November 2024; in Turkey the Union of Bar Associations' (TBB) guide of August 2026; in the UAE the Ministry of Justice's ten guiding principles for law firms (September 2026). None forbids AI. All attach conditions.

If a wrong citation is already in a filing

Speed matters, and so does the order of the steps.

1. Stop and re-check the whole document, not only the citation the judge found. Decisions in the database often involve several.
2. If it has not been filed or delivered, fix it now. The Spanish circular says that correcting in time can prevent the breach from being completed.
3. If it has been filed, tell the court yourself, in writing and plainly: what was wrong, what you have done. A prompt apology ended sanction proceedings in Navarra. In London the court referred even a solicitor who had admitted the error and reported himself, and lawyers who shrugged off five fake cases as "cosmetic errors" were told they should have been shocked.
4. Tell the client and your professional-indemnity insurer before they hear it elsewhere.
5. Ask whether your bar expects a self-report. The London judge said the lawyers should have reported themselves.
6. Write down how it happened and change the process. Courts and bars ask what you changed.

Where to verify

• United States: the court's own docket or PACER; CourtListener and Google Scholar for free searches; Westlaw, Lexis or Bloomberg Law for status checks.
• United Kingdom: BAILII and the National Archives' Find Case Law; legislation.gov.uk for statutes as amended.
• Canada, Australia: CanLII and AustLII.
• European Union: EUR-Lex for legislation and the Court of Justice; HUDOC for the European Court of Human Rights.

For research across many jurisdictions at once, see our guide to cross-border legal research.

Where Lexvora fits, and where it stops

Lexvora answers legal questions across many jurisdictions and, when its research runs, attaches the sources it found, so steps 1 and 4 start from a link you can open instead of a name you have to hunt for. A source link is a starting point, not proof: Lexvora is a research tool, it can be wrong like any other, and the six checks apply to its output too. If you want to see how that feels on a real question, you can try it free. Small practices will find more in why solo lawyers and small firms need AI legal tools.

Frequently asked questions

Is it unethical for a lawyer to use ChatGPT or another AI tool?

No guidance we reviewed forbids it. All of it attaches conditions: competence, confidentiality, supervision and, above all, verification before anything reaches a court or a client.

What happens if I file a fake citation by mistake?

The cases above run from a warning to a $5,000 sanction, fines of €420 to €1,800 in Spain, five minimum wages in Brazil, a wasted-costs order of AED 282,508 and referral to the bar or the regulator; a US federal court went to about $110,000 in one matter. Intent decides some consequences (contempt needs knowledge, and § 43a(3) of the German BRAO speaks of knowing untruths) but not others (a costs order does not). What you control is how fast and how candidly you correct.

Do I have to tell the court I used AI?

It depends on the court. Some require a declaration and many do not, so keep a table of the rules of the courts you appear in. Verification is necessary everywhere, but where a court requires a declaration it does not replace it.

Do legal AI tools such as Lexis+ AI or Westlaw hallucinate too?

Less often than general chatbots, but yes. Stanford researchers found hallucination rates between 17% and 33% across three such tools in 2024. Products change, so test the one you use on questions whose answers you already know.

How long does the six-step check take?

Plan on a few minutes per authority. Steps 1 and 2 are quick; the reading in steps 3 and 4 is the real work. For a brief with twenty authorities, plan on an hour, not a minute.

Can judges and experts hallucinate too?

The database lists 33 decisions in which a judge was the party concerned and 17 involving experts. Treat everything you rely on the same way, whoever produced it.

Sources and dates

Checked on 8 November 2026. Where reports differ, we followed the court's own text.

• AI Hallucination Cases Database, Damien Charlotin, updated 5 November 2026, under CC BY 4.0; we regrouped his figures (for example by lawyer) and computed the percentages. The July 2025 count is from Next, 1 July 2025.
• Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Magesh and others (Stanford RegLab and HAI), published in the Journal of Empirical Legal Studies in 2025.
• Ayinde v Haringey and Al-Haroun v Qatar National Bank, [2025] EWHC 1383 (Admin), the full judgment at the National Archives; Court issues stark warning to lawyers over AI-generated fake cases, Legal Futures, and the Law Gazette report on the first judgment.
• AG Köln, 312 F 130/25, beck-aktuell.
• Hallucinations before French courts, Morgan Lewis, March 2026.
• United States and Canada: ABA Journal on the $110,000 Oregon sanction and the Federal Court of Canada's page on AI.
• Spain: the database's Spanish entries (Galicia, and the February Canary Islands case), TSJ Navarra press note, TSJ Canarias and the CGAE circular.
• Brazil: TSE decision and OAB recommendations.
• Turkey: press report on the Ankara case, the lawyer's account of the court's order and the TBB guide.
• UAE: ADGM judgment analysis, Morgan Lewis, and the Ministry of Justice principles.

---

This article is for informational purposes only and does not constitute legal advice. Rules on AI in court filings differ by court and change quickly: check those that apply to you. Last reviewed: 8 November 2026.