AI Hallucinations in Court: The Sanctions Cases Every Lawyer Should Know, and How to Avoid Being Next

In June 2023 a federal judge in New York fined two lawyers and their firm $5,000 for filing a brief built on decisions that did not exist. The cases had been produced by ChatGPT, complete with plausible names, docket numbers and quotations. Mata v. Avianca became the cautionary tale of legal AI, and for a while many lawyers treated it as a one-off.

It was not. Since then courts have sanctioned lawyers at every level, from sole practitioners to attorneys at large firms, in trial courts, appellate courts and disciplinary tribunals. The penalties have grown sharply. This article sets out the leading decisions, the patterns that run through all of them, and a practical routine that keeps your name out of the next judgment.

What counts as an AI hallucination in a filing

A hallucination is output that reads as authoritative but is false. In legal filings it takes three main forms:

1. The invented authority. A case, statute or article that does not exist at all. This was the failure in Mata.
2. The misattributed proposition. A real case, correctly cited, offered for a point it does not decide.
3. The fabricated quotation. Words placed in quotation marks and attributed to a real judgment that never said them.

The third form is now the most dangerous, because a quick citation check passes. The case exists, the citation resolves, and only someone who reads the passage discovers the quotation is not there.

The leading cases, in order

The decisions below are a selection of the most cited. Fuller summaries, with links to each source, are on the AI Hallucination Sanctions Tracker.

Case	Court	Date	What happened	Outcome
Mata v. Avianca, Inc.	S.D.N.Y.	June 2023	Brief cited six nonexistent decisions generated by ChatGPT, with fabricated quotations	$5,000 sanction on lawyers and firm
People v. Crabill	Colorado Presiding Disciplinary Judge	November 2023	Motion relied on invented cases, and the lawyer misled the court about it	Suspension from practice
Park v. Kim	2d Cir.	January 2024	Reply brief cited a nonexistent case	Referral to the court’s Grievance Panel
Gauthier v. Goodyear	E.D. Tex.	November 2024	Nonexistent cases and fabricated quotations	$2,000 penalty and AI-related CLE
Wadsworth v. Walmart	D. Wyo.	February 2025	Eight nonexistent cases from a firm’s in-house AI tool	$5,000 total; one pro hac vice admission revoked
Lacey v. State Farm	C.D. Cal. (Special Master)	May 2025	Bogus citations across filings by two firms	$31,100 in sanctions
Ayinde v. Haringey	High Court of England and Wales	June 2025	Fictitious authorities in judicial review grounds	Regulator referral and a warning on contempt
Couvrette v. Wisnovsky	D. Or.	December 2025	15 nonexistent cases and 8 fabricated quotations; counsel refiled instead of withdrawing	$110,204.38 in sanctions and fees; case dismissed with prejudice
Whiting v. City of Athens	6th Cir.	March 2026	More than two dozen fabricated citations and misstatements of the record	$15,000 per attorney, fees, double costs, disciplinary referral
Nebraska Supreme Court divorce appeal	Neb.	April 2026	57 of 63 citations defective; the lawyer first attributed the brief to traditional research	Temporary licence suspension

Two features of the table matter more than any single figure. First, the penalties have climbed: a $5,000 fine in 2023 had become six figures and licence suspensions by 2026. Second, the courts are no longer confined to the United States. In Ayinde the Divisional Court, led by Dame Victoria Sharp, warned that freely available generative AI tools are not capable of reliable legal research and that more serious misuse could lead to contempt proceedings or a police referral.

The five patterns behind every case

Read the judgments together and the same failures repeat.

1. Nobody read the authority

In every reported case, at least one lawyer signed a filing without opening the cases it cited. Courts treat that as the core breach. The Sixth Circuit in Whiting put the duty in personal terms: the lawyer must read and verify every citation they put before the court.

2. Signing counsel assumed someone else checked

Wadsworth and Lacey involved more than one lawyer or firm. Each assumed another had verified the work. Courts sanction everyone whose name is on the filing. Supervision is not delegation of responsibility.

3. The tool was trusted because it was “built for lawyers”

Several reported cases involved legal research products rather than general chatbots. Purpose-built tools reduce error rates, but none of them guarantees accuracy. “The tool was designed for lawyers” has not persuaded any court.

4. The response to discovery made it worse

Compare the outcomes. Lawyers who promptly admitted the error, withdrew the filing and explained what happened generally received smaller penalties. Lawyers who denied using AI, blamed others or refiled the same material faced the heaviest sanctions. In Couvrette, refiling after notice was a central reason the total exceeded $110,000. In Crabill and the Nebraska matter, dishonesty about the source of the errors put the lawyer’s licence at risk.

5. The other side noticed first

In most cases it was opposing counsel, not the court, who found the problem. Your opponent will check your citations. So should you, and you should check theirs: in Withers v. City of Aberdeen (N.D. Miss., June 2026) lawyers on both sides were sanctioned.

Why AI tools invent law

Large language models generate text by predicting likely words. They do not look up a database of judgments unless the product is specifically connected to one, and even then the model can blend retrieved material with generated text. A citation in the right format, with a plausible party name and a realistic reporter volume, is exactly the kind of text these systems are good at producing. The model has no internal sense of whether the case exists.

That is why a hallucinated authority looks so convincing. It is fluent by design. Treat every AI-generated citation as a lead to be confirmed, never as a finding.

The verification routine that prevents it

The following routine takes minutes per filing and addresses every failure pattern above. It reflects the duties described in ABA Formal Opinion 512, which confirms that lawyers using generative AI remain responsible for competence, candour and the accuracy of what they file.

1. Pull every authority from a primary source. Open each case, statute and rule in an official reporter, court website or established legal database. If you cannot find it in two attempts, assume it does not exist.
2. Check each quotation word for word. Search the judgment for the quoted passage. If the words are not there, delete the quotation marks and paraphrase what the court actually said, or remove the point.
3. Read enough of the case to confirm the proposition. Confirm that the authority decides the point you cite it for and has not been reversed or overtaken.
4. Record who verified what. On team filings, keep a short log naming the person who checked each authority. It ends the “I thought you checked” problem.
5. If an error gets through, correct it at once. Tell the court and the other side, withdraw or correct the filing, and explain plainly. Every reported case suggests that candour limits the damage.

Use AI to find, organise and draft. Use your own eyes to verify. The signature on the filing is a personal certificate that you did.

What firms should put in place

Individual habits are not enough for a firm. A short written policy should require verification before filing, name the tools approved for client work, set rules for confidential data, and assign supervision duties for junior lawyers and staff. Training should use real examples from the cases above, because nothing changes behaviour faster than a judgment naming a lawyer who made the same assumption you are about to make.

If you are drafting that policy now, the article on what an AI use policy for a law firm must cover sets out the essential sections.

The bottom line

The lesson of three years of sanctions is consistent. Generative AI is a powerful drafting and research assistant, and courts have not prohibited it. What they will not accept is a lawyer who files work they have not checked. Verify every authority, verify every quotation, and correct mistakes candidly. Those three habits are the whole of the defence.