AI Agents for Lawyers

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.

By Dr. Niklas SchmidtUpdated 13 November 20269 min read
On this page
  1. What happened, in the order it happened
  2. The three failures Judge Castel actually punished
  3. Why "is this case real?" is the wrong question for a model
  4. How later courts cite and extend Mata
  5. What has changed since June 2023
  6. What has not changed
  7. What to do this week

Mata v. Avianca (S.D.N.Y., 22 June 2023) was the first widely reported court sanction for filing ChatGPT-invented case law: six fabricated opinions, a $5,000 joint penalty and a finding of subjective bad faith. Three years on, the decisive error was not using a chatbot but asking it to confirm its own citations, then standing by them after the court asked. That pattern still sets the penalty in 2026.

What happened, in the order it happened

Roberto Mata sued the airline Avianca over an injury from a serving cart on a flight, and Avianca moved to dismiss. The opposition brief, filed in March 2023 under the signature of Peter LoDuca but researched by his colleague Steven Schwartz, cited six decisions that do not exist, among them Varghese v. China Southern Airlines. Avianca's lawyers could not find the cases. Judge P. Kevin Castel ordered copies, and what came back were excerpts that were themselves ChatGPT output.

StageWhat the record shows
March 2023Opposition brief cites six non-existent decisions, with quotations
Spring 2023Avianca's counsel cannot locate the cases; the court orders copies
Spring 2023"Copies" filed, themselves ChatGPT excerpts; order to show cause issued
May 2023Schwartz's affidavit admits using ChatGPT and asking it whether the cases were real
8 June 2023Sanctions hearing before Judge Castel
22 June 2023Rule 11 sanctions; $5,000 jointly against both lawyers and the firm; letters to the client and to each judge wrongly identified as the author of a fabricated opinion; finding of "subjective bad faith"

The Wikipedia account records the $5,000 figure and the bad-faith finding. Against what followed, the money was trivial. The reasoning was not.

The three failures Judge Castel actually punished

Judge Castel did not hold that using AI was improper. Existing rules place a gatekeeping duty on the lawyer who signs, and that duty was abandoned at three separate points.

  1. Research in a tool that cannot research. ChatGPT in early 2023 had no legal database behind it and no way to retrieve a case. It composed text in the shape of a citation because that is what a language model does: it predicts plausible words.
  2. Asking the tool to check itself. Schwartz's affidavit described asking ChatGPT whether Varghese was a real case. It said yes, and that the case could be found on the major research databases. The answer was as fabricated as the case, and it was accepted without opening a database.
  3. Standing by the fakes after the court asked. The bad-faith finding rested on what happened after Avianca's counsel raised the alarm: instead of checking and withdrawing, the lawyers produced "copies" that compounded the problem. The order to show cause was the last clean exit, and it was not taken.

Later courts have weighed the third failure most heavily: the sanctions casebook shows fines for the first two errors and suspensions for the third.

Why "is this case real?" is the wrong question for a model

The instinct to ask the model is natural: it produced the citation, so it should know. It does not, and courts now treat the question as an aggravating factor rather than a defence.

  • OpenAI's own researchers concluded in "Why Language Models Hallucinate" (5 September 2025) that training and benchmarks reward guessing over "I don't know". A model asked to confirm its earlier answer is rewarded for confidence, not for retraction.
  • Stanford's 2026 AI Index reports that when a false premise is framed as the user's own belief, accuracy collapses: GPT-4o fell from 98.2% to 64.4%, DeepSeek R1 from above 90% to 14.4%. "Varghese is a real case, correct?" is exactly that framing.
  • Web search does not close the gap: a University of Pennsylvania study (3 April 2026) found search-augmented models and deep-research agents fabricate 3–13% of the URLs they cite, and agents hallucinate more because they cite more (arXiv).

The courts have caught up. On 27 April 2026 a New Jersey lawyer, Raja Rajan, was sanctioned a second time, $5,000, after "using AI to verify AI" (as reported by the ComplianceHub tracker). Damien Charlotin's database names "asking AI to self-validate" a decisive aggravator in sentencing.

One verification task a model does well is extracting every citation from a draft so a human can check each one. The prompt below asks for that and nothing else.

You are helping a lawyer prepare a citation check. Do NOT tell me whether any
authority is real, is good law or is correctly quoted. You cannot know that.

From the draft below, extract every authority cited (cases, statutes, rules,
regulations, secondary sources) into a table with these columns:

1. Citation exactly as written in the draft
2. Proposition the draft relies on it for (one sentence, quoting the draft)
3. Any quotation attributed to it (verbatim)
4. Where it appears (heading or paragraph number)

End with a row count. Do not add, correct, complete or "fix" any citation.
If a citation looks incomplete, copy it as written and flag it [INCOMPLETE].
If the same authority appears twice with different citations, list both and
flag [MISMATCH].

Draft:
[paste]

Every row then goes to Westlaw, Lexis, CourtListener, BAILII or the court's own site, and the lawyer reads the source. The citation-checking protocol covers the full routine, including misgrounded citations, where the case is real but says something else; Charlotin counted 854 of those by September 2026, and they are harder to catch than fabrications.

How later courts cite and extend Mata

Mata became the template because it separated the technology from the conduct, and every court since has applied the same logic under rules that predate ChatGPT. Norton Rose Fulbright's June 2026 review of the sanctions cases found that courts consistently rely on Rule 11, FRAP 38, 28 U.S.C. § 1927, inherent power and professional-conduct rules, with no AI-specific rule needed; candour mitigates and evasiveness aggravates.

CaseCourt, dateMata error repeatedOutcome
Park v. Kim2d Cir., 30 Jan 2024One ChatGPT-invented case in a briefReferral to the Grievance Panel
Wadsworth v. WalmartD. Wyo., 24 Feb 20258 of 9 cited cases fake, from Morgan & Morgan's in-house MX2.law tool$5,000 across three lawyers; pro hac vice revoked; Judge Rankin: "a fake opinion is not 'existing law'"
Johnson v. DunnN.D. Ala., 23 Jul 2025Butler Snow lawyers filed ChatGPT cases; the firm had an AI policy since June 2023Reprimand, removal from the case, bar referral; Judge Manasco: "If fines and public embarrassment were effective deterrents, there would not be so many cases to cite"
Couvrette v. WisnovskyD. Or., 12 Dec 2025Fabricated citations and quotations in three briefs; quiet refiling instead of disclosure$110,204.38 in total, the largest US penalty; case dismissed with prejudice
Whiting v. City of Athens6th Cir., 13 Mar 2026Over two dozen bad citations; lawyers called the show-cause order "void"$15,000 per lawyer plus fees and double costs
State v. LakeNeb. Sup. Ct., 15 Apr 202657 of 63 citations defective; AI use denied for weeksInterim suspension
Gamez v. County of FresnoE.D. Cal., Apr 2026Hallucinated citations, disclosed candidly by counsel who believed the tool had anti-hallucination protectionNo sanction

Read down the last column: the penalty tracks the response. Schwartz and LoDuca paid because they stood by the fakes; the Fresno lawyers did not and paid nothing. Anyone who drafts briefs and motions with AI should decide, before the first draft, what happens the day a citation is challenged.

What has changed since June 2023

June 2023September 2026
Known decisionsA handful2,041 in Charlotin's database (14 Sep 2026); 811 involving lawyers; roughly eight new decisions a day by June 2026
Court rulesJudge Brantley Starr (N.D. Tex.) issued the first AI certification order in May 2023Over 30 of 94 federal district courts require disclosure or certification, and 300+ individual judges have standing orders (April 2026)
Bar guidanceNone AI-specificABA Formal Opinion 512 (29 Jul 2024); 17 jurisdictions with formal opinions; California's May 2026 guidance: "No paper filed in court should contain citations that the attorney responsible for the filing has not personally read and verified"
Measured hallucination ratesGPT-3.5 69% and GPT-4 58% on legal questions in the 2024 "Large Legal Fictions" studyHAQQ benchmark (5 Jun 2026): GPT-5.5 3% citation hallucination, Claude Opus 4.8 about 12%, Mistral Large 64%; 24% of answers misapplied law that did not support the claim
Legal-grade toolsNot yet testedStanford RegLab measured Lexis+ AI at 17% and Westlaw AI-Assisted Research at 33% (2024 tools); sanctioned filings since involve CoCounsel, Westlaw Precision, vLex and Claude
Largest US penalty$5,000$110,204.38
Professional disciplineNoneInterim suspension (Nebraska), six-month suspensions (Ninth Circuit, M.D. Pa.), a dismissed appeal (Alabama), a strike-off in England (SRA v Kumar, August 2026)

Two rows deserve a second look: the best model's rate fell from 58% to 3% while penalties rose twenty-fold. A court does not grade on a curve, and at 3% a thirty-citation brief still more likely than not contains a fabrication.

What has not changed

The mechanism is the same: a plausible citation is indistinguishable from a real one until someone opens the source. Scientific American reported in May 2026 that lawyers keep citing fake cases because of automation bias and "cognitive surrender", the tendency to reverse a correct judgment when a confident machine disagrees. In Fletcher v. Experian (February 2026) the Fifth Circuit said the problem "shows no sign of abating". It now reaches the bench: in July 2025 Judges Neals (D.N.J.) and Wingate (S.D. Miss.) withdrew orders containing fabricated quotations after an intern used ChatGPT and a clerk used Perplexity, and Charlotin logged 32 judge entries by September 2026.

The reputational cost has grown. Morgan & Morgan's T. Michael Morgan, sanctioned in Wyoming in 2025, was denied pro hac vice admission in Massachusetts in May 2026 because the judge saw "no showing that he has learned from his mistakes". A Fortune op-ed by a LexisNexis executive that month asked whether any client would hire the lawyer who had just been sanctioned for AI use.

And the hallucination has changed shape rather than disappeared. The 2023 fakes were whole invented cases; the 2026 versions are subtler: real cases cited for propositions they do not support, quotations that paraphrase, a rule number one digit off. Whether AI can do legal research without hallucinating has an honest answer: not yet, and the errors that remain are the ones a skim will not catch.

What to do this week

The Mata lessons fit on one page and belong in the filing procedure, not in a training slide.

  1. Write the Mata rule down. California's formulation is the cleanest: no paper filed in court contains a citation the responsible lawyer has not personally read and verified. Put it in the firm AI policy and the pre-filing checklist.
  2. Ban self-verification. State expressly that asking any AI tool whether a citation is real or good law is not a verification step and counts as no check at all.
  3. Use the model for extraction only. Run the prompt above on every AI-assisted draft, check each row in a database and keep the completed table in the matter file as the audit trail.
  4. Know the local order. Check whether the court, and the individual judge, has an AI standing order before filing.
  5. Decide the confession protocol now. If a citation is challenged, the same-day answer is: check, disclose every affected citation, withdraw, explain. Gamez shows what candour buys; Couvrette shows what quiet refiling costs.
  6. Assume legal-grade tools fail too. Lacey v. State Farm ($31,100) involved CoCounsel and Westlaw Precision. Grounded retrieval lowers the rate; it does not remove the duty to read.

Module 6 of the AI Lab for Lawyers course works through Mata and its successors alongside the verification habits that keep a practice out of Charlotin's database. The guide to why hallucinations happen and how to catch them is the natural next read; the English counterpart is Ayinde v Haringey, where the Divisional Court reached the same conclusions two years later with a longer sanctions list.

Frequently asked questions

What happened in Mata v. Avianca?

In a personal-injury claim against the airline Avianca, the plaintiff's lawyers filed a March 2023 brief citing six decisions that ChatGPT had invented. When Avianca's counsel could not find them, the lawyers filed ChatGPT-generated excerpts as copies. On 22 June 2023 Judge Castel found subjective bad faith, imposed a $5,000 joint sanction and ordered letters to the client and to the judges wrongly named as authors.

Why was asking ChatGPT to verify its own cases a mistake?

A language model predicts the most plausible next words; asked whether a citation it produced is real, the plausible answer is yes. Research published by OpenAI in September 2025 shows training rewards confident guessing over admitting ignorance, and Stanford's 2026 AI Index shows accuracy collapses when a user frames a false premise as their own belief. Only a legal database can verify a citation.

Is Mata v. Avianca still relevant in 2026?

Yes. Courts still cite it as the template, and the same three errors recur. Damien Charlotin's database listed 2,041 decisions involving hallucinated material on 14 September 2026, 811 of them involving lawyers, with the largest US penalty at $110,204.38 in Couvrette v. Wisnovsky and suspensions in Nebraska, the Ninth Circuit and Pennsylvania during 2026.

What sanctions did the Mata lawyers receive?

Judge Castel imposed a $5,000 penalty jointly on Steven Schwartz, Peter LoDuca and their firm under Rule 11, and required them to notify their client and each judge falsely identified as the author of a fabricated opinion. By the standards of 2026, when the Sixth Circuit fined two lawyers $15,000 each and the Nebraska Supreme Court suspended a lawyer on an interim basis, that was lenient.

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

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

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

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