AI Agents for Lawyers

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.

By Dr. Niklas SchmidtUpdated 13 November 202612 min read
On this page
  1. The numbers behind the casebook
  2. 2023 to 2024: the warnings
  3. 2025: fines reach Big Law and cross the Atlantic
  4. 2026: suspensions, dismissals and a strike-off
  5. Six patterns from the casebook
  6. Judges are in the database too
  7. Why the warnings keep failing
  8. The protocol that would have stopped every case
  9. Where this leaves you

Damien Charlotin's AI Hallucination Cases database listed 2,041 court decisions worldwide on 14 September 2026, 811 of them involving lawyers. The penalties have climbed from $5,000 in Mata v. Avianca (S.D.N.Y. 2023) to $110,204.38 in Couvrette v. Wisnovsky (D. Or. 2025), and in 2026 to suspensions, a dismissed appeal and a strike-off. One pattern runs through all of them: the response to the hallucination sets the penalty, not the hallucination itself.

The numbers behind the casebook

Charlotin, a research fellow at HEC Paris, logs a decision only when a court found or clearly implied that a party relied on hallucinated material; allegations are excluded, which is why he calls the count "necessarily an undercount". The database is the reference for everything below.

DateDecisions in the database
Mid-2025About 200
29 Dec 2025712 (90% of them from 2025)
Early Apr 20261,227
9 Jun 20261,598 (about eight new cases a day)
14 Sep 20262,041

By country: USA 1,395; Canada 217; Australia 111; UK 69; Israel 57; Brazil 41; Germany 11. By actor: litigants in person 1,174; lawyers 811; judges 32; experts 15. By nature: fabricated case law 1,690 instances; misrepresented real authority 854; false quotations 548. Scientific American reported in May 2026 that growth had plateaued at 350 to 400 decisions a quarter, and US courts imposed over $145,000 in penalties in the first quarter of 2026 alone.

2023 to 2024: the warnings

Mata v. Avianca (S.D.N.Y., 22 June 2023) set the template: six fabricated cases, a lawyer who asked ChatGPT whether they were real and accepted its answer, a finding of "subjective bad faith" by Judge Castel and $5,000 jointly. The Mata page covers what it still teaches. Park v. Kim (2d Cir., 30 January 2024) referred a lawyer to the grievance panel for a ChatGPT-invented case, and People v. Crabill (Colorado, November 2023) produced the first bar discipline, a suspension of one year and one day. The technology was new, the duty was old, and courts said so in almost every order.

2025: fines reach Big Law and cross the Atlantic

CaseCourt, dateWhat went wrongSanction
Kohls v. EllisonD. Minn., Jan 2025A Stanford professor's expert declaration on deepfakes cited AI-invented studiesDeclaration struck
Wadsworth v. WalmartD. Wyo., 24 Feb 2025Morgan & Morgan; 8 of 9 cited cases fake, generated by the firm's in-house MX2.law tool$5,000 across three lawyers; pro hac vice revoked; Judge Rankin: "a fake opinion is not 'existing law'"
Lacey v. State FarmC.D. Cal., 6 May 2025K&L Gates and Ellis George; 9 of 27 citations wrong; CoCounsel, Westlaw Precision and Gemini used$31,100
Concord v. AnthropicN.D. Cal., May 2025Latham & Watkins asked Claude to format a citation; it invented the title and authors; counsel apologised 15 May 2025Part of the declaration struck; Judge van Keulen: "a world of a difference between a missed citation and hallucination generated by AI"
Ayinde v Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin)Divisional Court, 6 Jun 2025Five fake authorities in one case, 18 of 45 in the other; a barrister's denial found "unpersuasive"£2,000 plus VAT wasted costs each; BSB and SRA referrals; a sanctions ladder up to contempt and police referral
Bandla v SRA [2025] EWHC 1167 (Admin)High Court, May 2025Fictitious authorities cited against the solicitors' regulator itselfFabrications placed on the record; cited as a precursor to Ayinde
AG Köln 312 F 130/25Cologne, 2 Jul 2025Invented BGH decisions, commentary paragraphs and journal articles, "frei erfunden"Warning of a § 43a Abs 3 BRAO breach
Johnson v. DunnN.D. Ala., 23 Jul 2025Three Butler Snow lawyers; ChatGPT-invented cases; the firm had an AI policy since June 2023Public reprimand, removal from the case, Alabama State Bar referral
Coomer v. LindellD. Colo., Jul 2025Mike Lindell's lawyers filed fabricated citations$3,000 per lawyer
Noland v. Land of the FreeCal. Ct. App., 12 Sep 2025"Nearly every" quotation fabricated; California's first published opinion$10,000; State Bar referral; opposing counsel criticised for missing the fakes
OGH 14 Os 95/25iAustrian Supreme Court, 7 Oct 2025A Graz defence lawyer's AI-drafted appeal cited invented OGH decisionsDismissed without reaching the merits; § 9 Abs 1 RAO
Kammergericht Berlin 17 WF 144/25Berlin, 20 Nov 2025Non-existent BGH decision from a "fantasierende KI"Rebuke; Prüfpflicht restated
Ko v. Li, 2025 ONSC 6785Ontario, 4 Dec 2025Lawyer blamed staff for a ChatGPT-tainted factum, then confessed to lying about who drafted itReferred for criminal contempt
Couvrette v. WisnovskyD. Or., 12 Dec 202515 non-existent cases and 8 fabricated quotations across three briefs; counsel quietly refiled rather than confess$110,204.38: about $80,000 in fees plus fines of roughly $15,000 and $14,000; case dismissed with prejudice

Couvrette is the record, and the ABA Journal account is worth reading in full. Magistrate Judge Mark D. Clarke wrote: "If there was ever an 'appropriate case' to grant terminating sanctions for the misuse of artificial intelligence, this is it." The plaintiff lost her $12 million winery dispute on her lawyers' citations, not on the facts.

2026: suspensions, dismissals and a strike-off

CaseCourt, dateWhat went wrongSanction
Fletcher v. Experian5th Cir., Feb 202616 fabricated quotations; vLex and CoCounsel involved; the problem "shows no sign of abating"$2,500
Whiting v. City of Athens6th Cir., 13 Mar 2026More than two dozen bad citations; counsel refused to answer the court's seven questions and called the show-cause order "void"$15,000 per lawyer plus fees and double costs
Oregon Court of AppealsMar 202615 false citations and 9 fake quotations$10,000; Chief Judge Lagesen ordered staff to log time spent on fabricated authority
Sullivan & CromwellBankr. S.D.N.Y., Apr 2026Fake citations in an 8 April 2026 filing, flagged by Boies SchillerApology to Chief Judge Martin Glenn; no fine
Nebraska Supreme Court15 Apr 202657 of 63 citations in a divorce appeal defective; the lawyer denied AI use for weeksInterim licence suspension, per WOWT
Alabama Supreme Court24 Apr 2026"Egregious" AI-tainted briefs; the apology paragraph cited two more non-existent casesAppeal dismissed; $17,200 in fees; bar referral; no filing without a co-signing attorney, per Bloomberg Law
Gamez v. County of FresnoE.D. Cal., Apr 2026Fabricated citations; counsel was candid and believed the tool had "anti-hallucination" protectionNo sanction
Coomer v. Lindell (second order)D. Colo., May 2026Repeat offence by the same lawyer$5,000
Lnu v. Blanche9th Cir., 3 Jun 2026Fabricated citations first called "typographical errors"Six-month suspension before the court and $2,500 each, per Bloomberg Government
Cork v Smith [2026] EWHC 1199 (Ch)Chancery Division, 4 Jun 2026A Pinsent Masons associate filed an invented "Insolvency Rule 12.37(5)" after the tool had flagged possible inaccuracyFirm referred to the SRA; extra costs borne by the firm
SRA v KumarSDT, 25 Aug 2026A registered foreign lawyer defended himself with AI-fabricated quotations, then "corrected" them with moreStruck off the Register of Foreign Lawyers
Booker v. KrogerN.D. Ga., 28 Aug 2026Fabricated citations$8,000 plus CLE
New Mexico Supreme CourtAug to Sep 2026A murder-appeal brief cited police testimony and witnesses ChatGPT had invented$5,000, removal and a bar referral; Justice C. Shannon Bacon: "Counsel, do you watch the news?"
Beus Gilbert PLLC v. BYUD. Utah, 9 Sep 2026Fabricated citations$3,000 plus two AI-ethics CLE courses

Norton Rose Fulbright's June 2026 review draws the doctrinal conclusion: courts have consistently held that existing rules suffice, namely Rule 11, FRAP 38, 28 U.S.C. § 1927, inherent power and professional-conduct rules. No AI-specific rule is needed to sanction, and none is coming to protect you.

Six patterns from the casebook

The cover-up costs more than the hallucination. Couvrette's record penalty was for quietly refiling. Nebraska's suspension followed weeks of denial. Whiting's $15,000 per lawyer followed a refusal to answer the court. Ko v. Li became a contempt referral when a false explanation collapsed. Kumar was struck off for correcting fabrications with fabrications. Against those, Gamez produced no sanction for a candid lawyer, and Morgan & Morgan's remediation in Wadsworth kept the total to $5,000. Candour does not always save you, as U.S. v. Farris (6th Cir., March 2026) showed when a lawyer with a 40-year clean record was removed from the CJA panel, but its absence never helps.

Legal-grade and in-house tools do not immunise. Wadsworth used a firm-built tool. Lacey used CoCounsel, Westlaw Precision and Gemini. Fletcher used vLex and CoCounsel. Concord used Claude in the hands of one of the world's largest firms. Grounded platforms reduce fabricated cases but still produce misgrounded ones, a real authority cited for a proposition it does not support, which is the category Charlotin logs 854 times and which Clearbrief's founder has noted is "not easy to spot with the human eye". The hallucination-rate comparison explains why no tool reaches the zero courts require.

A policy is not a defence. Butler Snow adopted an AI policy in June 2023 and three of its lawyers were reprimanded and removed in July 2025. Gordon Rees was reportedly sanctioned twice, in October 2025 and February 2026, despite updated policies. Judge Anna Manasco in Johnson v. Dunn explained why fines alone have failed: "If fines and public embarrassment were effective deterrents, there would not be so many cases to cite."

AI cannot verify AI. Mata's lawyer asked ChatGPT whether its cases were real. In April 2026 a New Jersey lawyer was reportedly sanctioned a second time, $5,000, for "using AI to verify AI". Charlotin's tracker names self-validation as a decisive aggravating factor. Verification means a database and a human; the verification protocol sets out the sequence.

Courts now prescribe education and publish a tariff. Beus Gilbert and Booker added CLE courses to fines. Oregon courts price fabrications at $500 per fake citation and $1,000 per fake quotation, a tariff that would have produced $16,500 in the Court of Appeals case where the court imposed $10,000. A California attorney accepted a 30-day suspension plus ten hours of technology CLE from the State Bar.

The other side's fakes are your problem too. Noland criticised respondent's counsel for failing to notice the appellant's fabricated citations. With litigants in person filing machine-written bundles, including a 300-page skeleton "generated entirely by ChatGPT" reported by the Law Gazette on 18 September 2026, checking the opponent's authorities is now part of preparation.

Judges are in the database too

Thirty-two entries involve the court's own ruling. Judge Julien Neals (D.N.J.) and Judge Henry Wingate (S.D. Miss.) withdrew orders in July 2025 containing fabricated quotations and non-existent parties; in October 2025 letters to Senator Grassley they admitted an intern had used ChatGPT and a clerk had used Perplexity. On 31 August 2026 a Fifth Circuit panel, with Judge Jerry Smith calling it "a pretty serious matter", heard Mississippi's request to reassign Judge Wingate's case because hallucinated citations survived into the amended order, as the ABA Journal reported. A survey of 112 federal judges published in March 2026 found 61.6% use at least one AI tool and 45.5% had received no court-provided training. The English judiciary's guidance of 31 October 2025, which makes judges responsible for all material in their name, is the institutional answer; the UK guidance page covers it.

Why the warnings keep failing

Scientific American's May 2026 analysis attributes the persistence to automation bias, time pressure and "cognitive surrender": people who have made a correct judgment reverse it when a machine disagrees, and awareness campaigns alone barely move verification rates under deadline. The Stanford HAI AI Index 2026 adds a mechanism: hallucination rates across 26 leading models ranged from 22% to 94% on a new factuality benchmark, and when a false premise was framed as the user's own belief, one model's accuracy fell from 98.2% to 64.4%. A lawyer who tells the model what the law is will get a citation for it. Web search does not close the gap: a University of Pennsylvania study in April 2026 found deep-research agents fabricating 3% to 13% of URLs, and the HAQQ benchmark of June 2026 found 24% of answers citing or applying law that did not support the claim. The hallucinations guide explains the mechanics; the fix is procedural.

The protocol that would have stopped every case

  1. Treat every citation, quotation and rule reference from any AI tool, including grounded legal platforms, as unverified until a human opens the source.
  2. Open the authority in a database's ordinary search, not its AI assistant. Confirm the case exists, the court and year match, and the pinpoint paragraph says what is claimed.
  3. Check every quotation word for word against the reported text; misquotation of a real case survives a quick existence check.
  4. Never ask the model whether its own output is correct, and never ask a second model either.
  5. Record who checked what and when in the matter file, in a form that can be produced to a court within the hour.
  6. Have the signing lawyer, not the drafter, read every authority before signing; sanctions attach to the signature under many standing orders.
  7. If an error is found after filing, write to the court and the other side the same day, identify every affected citation and explain what changed.
  8. Check the opponent's authorities with the same rigour before every hearing.

The brief and motion drafting guide builds this into a drafting workflow, and the litigation guide shows where AI still earns its place once verification is routine.

Where this leaves you

The cases are not an argument against using AI; they are an argument about procedure. Every sanction in the tables above followed a missing step that takes minutes, and every escalation followed a decision to conceal rather than correct. Malpractice cover is silent for now, with no major US carrier having filed an AI exclusion as of 5 May 2026, but application forms already ask about governance. Module 6 of the AI Lab for Lawyers course works through Mata, Couvrette, Ayinde and Cork v Smith in one session and ends with the verification drill above run live on a real brief. The Ayinde explainer is the place to start for the English cases, and the glossary entry gives clients the one-paragraph answer to why their lawyer now reads every case twice.

Frequently asked questions

How many lawyers have been sanctioned for AI hallucinations?

Damien Charlotin's AI Hallucination Cases database recorded 2,041 decisions worldwide on 14 September 2026, of which 811 involved lawyers; the rest are mostly litigants in person, plus 32 judges and 15 experts. Not every entry is a sanction, but the database only counts cases where a court found or clearly implied reliance on hallucinated material.

What is the biggest fine for AI fake citations so far?

In Couvrette v. Wisnovsky (D. Or.), Magistrate Judge Mark D. Clarke imposed $110,204.38 in total in December 2025, mostly the defendants' fees, after three briefs contained fabricated citations and invented quotations; the case was dismissed with prejudice. The Sixth Circuit's $15,000 per lawyer in Whiting v. City of Athens (March 2026) is the largest appellate fine.

Can I be sanctioned if I used Westlaw or CoCounsel rather than ChatGPT?

Yes. Lacey v. State Farm involved CoCounsel, Westlaw Precision and Gemini and cost $31,100. Fletcher v. Experian involved vLex and CoCounsel. Wadsworth v. Walmart involved a firm's in-house tool. Courts sanction the failure to verify, not the choice of tool, and grounded legal platforms still produce misgrounded citations.

What should I do if I discover a hallucinated citation after filing?

Tell the court and opposing counsel the same day, identify every affected citation, file a corrected document and explain what went wrong and what you have changed. In Gamez v. County of Fresno candid counsel received no sanction. Every escalated case, from Nebraska to Ko v. Li in Ontario, followed a denial, a quiet refiling or a false explanation.

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.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

Regulation & Ethics

Mata v. Avianca, Three Years On: What the First AI Sanction Still Teaches

Mata v. Avianca (S.D.N.Y. 2023) cost two lawyers $5,000 for six fake ChatGPT cases. What went wrong, why asking AI to verify AI fails, and what changed since.

9 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

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

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

Use Cases

Drafting Briefs and Motions with AI Without Ending Up in Charlotin's Database

AI brief drafting is safe only as a five-stage workflow: grounded research, outline, drafting from verified authorities, every case read, court rules checked.

10 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