AI Hallucination Sanctions: Lawyer Cases and How to Avoid Them

AI Hallucination Sanctions, in One Line

Courts sanction lawyers when they file briefs that cite cases, quotations or holdings an AI tool invented and nobody checked. The amounts and penalties vary a lot. In Mata v. Avianca (2023) two lawyers and their firm paid a $5,000 penalty. In Couvrette v. Wisnovsky (D. Or.) the orders of 12 December 2025 and 23 March 2026 add up to $110,204.38 across two lawyers, which is a $15,500 penalty plus the other side's fees, and the court dismissed the claims. Appeals courts have joined in: the Fifth, Sixth and Ninth Circuits each issued sanctions orders in 2026. The fix is not avoiding AI. It is checking every citation and every quotation against a primary source before you file.

Reviewed and corrected on 12 October 2026. Every case below was re-read against the court's own order. This is general information, not legal advice; pull the order before you rely on it.

TL;DR

• Lawyers keep getting sanctioned for filing unchecked AI output. The orders below range from a $1,000 fine for a signing lawyer in Wadsworth v. Walmart to a one-year-and-a-day suspension in Colorado.
• The money varies by case, not by formula. Examples from the orders: $5,000 in Mata, $3,000 for the drafter in Wadsworth, $10,000 in Noland, $15,000 per lawyer in Whiting, $31,100 in Lacey, and $110,204.38 in Couvrette. Several of those totals include fee-shifting.
• Appeals courts are acting. The Sixth Circuit fined two lawyers $15,000 each in Whiting (13 March 2026), the Fifth Circuit fined one lawyer $2,500 in Fletcher v. Experian (18 February 2026), and the Ninth Circuit fined two lawyers $2,500 each and suspended them from practice before it for six months in Lnu v. Blanche (3 June 2026).
• Real cases with invented quotations count too. In Farris (6th Cir.) the court said no case was fake, yet it still denied counsel his fee and referred him for discipline.
• Several courts said the lawyer's response mattered. The Ninth Circuit tied its discipline to "this repeated failure of candor." The Fifth Circuit wrote that "it is likely that the court would have imposed lesser sanctions" had counsel been forthcoming.
• ABA Formal Opinion 512 (29 July 2024) applies the existing Model Rules to generative AI. It does not say how much verification is enough: the level "will necessarily depend on the GAI tool and the specific task."

Part of our legal AI verification and hallucination guide series.

For related coverage, see the AI hallucination sanctions tracker, How to Verify AI Legal Citations Before You File and Damien Charlotin's Hallucination Tracker, Read Like a Risk Manager.

Where it started: Mata v. Avianca

On 22 June 2023, Judge P. Kevin Castel of the Southern District of New York sanctioned lawyers Peter LoDuca and Steven Schwartz and their firm, Levidow, Levidow & Oberman, in Mata v. Avianca, Inc. (No. 1:22-cv-01461). The order imposed a $5,000 penalty, jointly and severally, "pursuant to Rule 11, or, alternatively, its inherent authority." It also required letters to the plaintiff and to each of the six judges falsely named as authors of the fake opinions, among them the "Varghese" decision.

The opposition brief cited non-existent opinions that ChatGPT had produced. Avianca's counsel listed seven decisions it could not find. Later, Schwartz filed screenshots of his questions to the same chatbot, including whether Varghese was real. ChatGPT answered that the case existed and could be found on Westlaw and LexisNexis.

Judge Castel did not fault the technology itself. He wrote that there is "nothing inherently improper about using a reliable artificial intelligence tool for assistance," but that "existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings." It found bad faith on the part of the two individual lawyers "based upon acts of conscious avoidance and false and misleading statements to the Court."

The orders since then

Here is what the courts actually wrote, in rough order of date. Each link goes to the order or opinion.

Colorado, People v. Crabill (22 November 2023). Zachariah Crabill had never drafted a motion to set aside a judgment before. He cited cases he found through ChatGPT and did not read or verify them. Before the hearing he discovered the cases were incorrect or fictitious, but he did not tell the court or withdraw the motion. When the judge raised concerns he falsely blamed a legal intern. Six days after the hearing he filed an affidavit saying he had used ChatGPT. The Presiding Disciplinary Judge approved a stipulation suspending him for one year and one day, with 90 days to be served and the rest stayed on successful completion of a two-year probation. The discipline rested on Colorado Rules 1.1, 1.3, 3.3(a)(1) and 8.4(c) (approved summary).

Wyoming, Wadsworth v. Walmart (24 February 2025). Plaintiffs' motions in limine cited nine cases, and eight did not exist. Lawyer Rudwin Ayala had uploaded the draft to his firm's in-house platform, MX2.law, and asked it to add federal case law from Wyoming on motions in limine. Judge Kelly H. Rankin revoked his pro hac vice admission and fined him $3,000, noting his "honesty and candor." T. Michael Morgan and Taly Goody, who signed the motions and apparently were not given them to review beforehand, were fined $1,000 each (order).

C.D. Cal., Lacey v. State Farm (5 May 2025). A special master, retired Magistrate Judge Michael R. Wilner, found that the lawyers at two firms "collectively acted in a manner that was tantamount to bad faith." A lawyer at one firm used AI tools to generate an outline. Lawyers at the other firm built it into a brief and apparently nobody cite-checked it. The briefs were struck and the two firms were ordered to pay $31,100 jointly and severally. He called it "a collective debacle" and declined to sanction the individual lawyers further (order).

California, Noland v. Land of the Free (12 September 2025). In an opinion certified for publication, the Second District Court of Appeal found that 21 of 23 case quotations in the opening brief were fabricated. Counsel Amir Mostafavi told the court he had written drafts, "enhanced" them with ChatGPT and not read the result. The court ordered him to pay $10,000 and sent the opinion to the State Bar (opinion).

D. Or., Couvrette v. Wisnovsky (12 December 2025 and 23 March 2026). Three summary-judgment briefs cited fifteen non-existent cases and fabricated quotations attributed to eight real authorities. Magistrate Judge Mark D. Clarke wrote that "this case is a notorious outlier in both degree and volume." The December order struck the briefs, dismissed the plaintiffs' claims with prejudice and ordered Stephen Brigandi, the lawyer admitted pro hac vice, to pay $15,500 to the court. It also awarded the defendants their fees. The March order set those fees at $94,704.38: Brigandi pays $80,498.72 and local counsel Timothy Murphy pays $14,205.66, which the court imposed because he failed to meaningfully participate in the case, not because he signed the briefs. The total is $110,204.38 (December order, March order). We break it down in the Oregon $110K sanction piece.

5th Cir., Fletcher v. Experian (18 February 2026). The court found that counsel used generative AI to draft a substantial portion of her reply brief, if not all of it, and did not check it, and fined her $2,500. The reply had 16 fabricated quotations and five other serious misstatements. Counsel named Casetext (now part of Thomson Reuters' CoCounsel) and vLex among her sources, and the court noted both are generative AI products. It did not find which tool produced the errors (opinion).

6th Cir., Whiting v. City of Athens (13 March 2026). The court found "over two dozen fake citations and misrepresentations of fact" across three consolidated appeals. It asked counsel whether generative AI was used and they did not answer, so the opinion names no tool. It ordered full appellate fees and double costs under Appellate Rule 38, plus $15,000 from each lawyer as punitive sanctions under its inherent authority, because "smaller fines have plainly been inadequate." It cited aggravating factors: one lawyer's 2017 public censure for lack of candor, the other's five-year suspension from the Eastern District of Tennessee in August 2025, and their defiance of the show-cause order (opinion).

6th Cir., United States v. Farris (3 April 2026). Court-appointed counsel admitted using Westlaw's CoCounsel to draft both briefs. They carried invented quotations and misdescribed the holdings of real cases. The court denied payment under the Criminal Justice Act for the appeal and referred the matter for possible discipline. It said the risk remains "even when new tools are sponsored by trusted legal technology providers" (opinion).

9th Cir., Lnu v. Blanche (3 June 2026). The court found "multiple nonexistent cases, misattributed quotations, and gross misrepresentations of real cases" in immigration briefs. At the firm, law school graduates who were not yet licensed wrote the legal arguments and no attorney vetted their citations. The lawyers first called the errors typographical, and at argument one of them repeatedly said AI "was not used" before conceding it was "possible." The court fined Mike Singh Sethi and William Rounds $2,500 each, suspended them from practice before the Ninth Circuit for six months, and required a sworn statement about AI use and cite-checking in all their future filings for two years. It wrote that "lawyers using generative AI must thus be aware of the tendency of generative AI to make these mistakes and guard against them" (opinion).

For more rows, including Nebraska, the Fourth Circuit and Mississippi, see the sanctions tracker and the Nebraska case.

What the orders say about the response

In several of these orders the judges linked the penalty to the lawyer's behavior once the problem came to light. No formula comes out of that, since the totals also reflect the number of fakes and any fee-shifting.

• Mata: the bad-faith finding rested on "conscious avoidance and false and misleading statements," including shifting accounts of how the cases were found.
• Couvrette: deleting fake cases but leaving the statements of law was no correction, and the court called it a coverup.
• Fletcher: the lawyer's first response did not mention AI. The court said that had she accepted responsibility, lesser sanctions were likely.
• Lnu: the court said that had the lawyers disclosed AI use and apologized in their Motion to Correct, lesser sanctions may have been warranted.
• Crabill: the stipulated findings include that he falsely blamed an intern.

How big is the problem?

Damien Charlotin, a French lawyer and data scientist (the Fifth Circuit's description), keeps the public AI Hallucination Cases database. On 12 October 2026 its page said it was last updated on 10 October 2026 and listed 2,184 cases worldwide, 1,492 of them in the United States. It counts decisions in which a court addressed hallucinated content, including warnings, so it is wider than a sanctions list. It also says it does not track the whole universe of fake citations. The Fifth Circuit counted 239 US lawyer cases in the database when it ruled in February 2026. See our read of the database.

Stanford researchers measured the problem in two studies. In "Large Legal Fictions" (2024) they found that general-purpose models hallucinated between 58% (ChatGPT 4) and 88% (Llama 2) of the time on specific, verifiable questions about random federal court cases. In "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools" (Journal of Empirical Legal Studies, 2025), they tested tools from LexisNexis and Thomson Reuters. Those tools hallucinated less than GPT-4 but still "each hallucinate between 17% and 33% of the time." The study found that provider claims such as "hallucination-free" were overstated. It was run on 2024 product versions, so current tools may differ.

Some of the errors are subtle. The Stanford authors count a response as a hallucination if it is false, or if the citation is real but does not support the claim. That matches what the courts found in Farris, Noland and Lnu: real cases with invented quotations or misdescribed holdings.

Why legal AI hallucinates

A language model predicts likely text. It has no built-in list of cases and no concept of truth. Asked for a case supporting a point, it can produce something shaped like a citation, with plausible party names and a plausible reporter.

Retrieval-augmented generation (RAG) adds a search step before the model writes, so the answer is built from retrieved documents. That reduces errors, but the Stanford study shows it does not remove them. Retrieval can miss the best authority. The model can mischaracterize what it retrieved. And fabricated and accurate text look equally confident.

What ABA Formal Opinion 512 says

On 29 July 2024 the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512 on generative AI tools. It applies existing Model Rules rather than creating new ones. A fuller walk-through is in our ABA 512 guide. The points that matter for hallucinations:

• Competence (Rule 1.1). Lawyers "need not become GAI experts," but must have "a reasonable understanding of the capabilities and limitations of the specific GAI technology that the lawyer might use." Relying on or submitting a tool's output without "an appropriate degree of independent verification or review" could violate Rule 1.1. The right amount of review "will necessarily depend on the GAI tool and the specific task that it performs."
• Candor (Rules 3.1, 3.3 and 8.4(c)). Output used in court must be "carefully reviewed to ensure that the assertions made to the court are not false." The opinion notes that "even an unintentional misstatement to a court can involve a misrepresentation under Rule 8.4(c)."
• Confidentiality (Rule 1.6). Because many self-learning tools are designed so that output could lead to disclosure, informed consent is required before inputting information relating to the representation into such a tool, and "merely adding general, boiler-plate provisions to engagement letters" is not enough.
• Supervision (Rules 5.1 and 5.3). Managerial lawyers "must establish clear policies regarding the law firm's permissible use of GAI."
• Fees (Rule 1.5). Lawyers who bill by the hour must bill the time they actually spend.

512 is guidance, not a rule, and it does not cite Mata. The sanctions come from Rule 11, appellate rules, inherent authority and state discipline systems.

Court rules and bar guidance differ by court

The orders above rest on rules that predate ChatGPT. Federal Rule of Civil Procedure 11(b) makes presenting a paper a certification that the legal contentions are warranted by existing law after a reasonable inquiry. Appellate courts used Appellate Rule 38 (Whiting), Rule 46 (Fletcher, Lnu) and inherent authority.

Some courts have gone further. The Fifth Circuit says it considered a rule requiring a certification about generative AI and declined to adopt it, concluding that existing rules were sufficient. In Lnu the Ninth Circuit imposed a two-year certification requirement on the sanctioned firm. Because requirements differ by court and judge, check the local rules and the standing orders where you file. Our state-by-state AI rules guide tracks bar opinions and court rules.

A practical verification routine

Treat this as a checklist drawn from the orders, not a legal rule. Our full checklist goes deeper.

1. Treat every AI-supplied citation as a lead, not a source. Mata shows a chatbot assuring a lawyer a fake case was real.
2. Open each case in a primary source. Westlaw, Lexis, Google Scholar, PACER, CourtListener, Justia or the court's own site all work. If it does not resolve, it is not a case.
3. Check each quotation word for word. In Farris the cited cases existed but the quotations did not, and in Noland the court said most of the cases the fabricated quotes were attributed to exist, so an existence check alone would have passed many of them.
4. Read the case and confirm it supports your proposition. The Ninth Circuit in Lnu said a competent lawyer must read the cases cited.
5. Check that the authority is still good law. Run a citator.
6. Check statutes and rules against the current official text. AI tools can misquote statutes or cite repealed provisions.
7. Supervise nonlawyers and juniors. Farris faulted reliance on staff, and Lnu involved unlicensed brief writers whose citations no lawyer vetted.
8. Have a response plan. If a cite is flagged, verify it in a primary source, tell the court and opposing counsel quickly, and fix the proposition as well as the citation. Lnu states the instruction: "immediately alert the court and opposing counsel of the error and disclose its source."
9. Keep a record of what you checked and how, especially where a court requires AI disclosure.

What to do next

• Audit your tools. Know which AI tools your lawyers and staff actually use.
• Write a policy. 512 expects managerial lawyers to set clear policies on permissible AI use. Name approved tools and require citation verification.
• Train people. 512 says supervisors must make sure lawyers and nonlawyers understand the tools and their risks.
• Track local rules in every court where you file.

FAQ

Can a lawyer be sanctioned for using AI?

In the orders above the courts sanctioned filing unverified output, not using AI. The Ninth Circuit wrote in Lnu: "We do not sanction Sethi and Rounds for the simple fact that they or their subordinates used generative AI." Rule 11 and ABA Formal Opinion 512 leave the signing lawyer responsible for what is filed.

What was the Mata v. Avianca case?

Mata v. Avianca, Inc. (No. 1:22-cv-01461, S.D.N.Y., 22 June 2023) was the first high-profile AI hallucination sanction. Lawyers Steven Schwartz and Peter LoDuca and their firm filed an opposition that cited fake opinions ChatGPT produced. Judge Castel imposed a $5,000 penalty, jointly and severally, and found bad faith on the part of the two lawyers.

What is the largest AI hallucination sanction?

The largest total we found in a US court order is $110,204.38 in Couvrette v. Wisnovsky (D. Or.): a $15,500 penalty on Stephen Brigandi plus $94,704.38 in the defendants' fees, split between him and local counsel Timothy Murphy. As of Charlotin's 10 October 2026 update, we found it to be the largest US monetary entry among lawyer-involved cases in the database's CSV.

Have federal appeals courts sanctioned lawyers over AI?

Yes. In 2026 the Fifth Circuit fined one lawyer $2,500 (Fletcher v. Experian), the Sixth Circuit fined two lawyers $15,000 each plus fees and double costs (Whiting v. City of Athens), and the Ninth Circuit fined two lawyers $2,500 each and suspended them from the court for six months (Lnu v. Blanche).

Can I be sanctioned if I only signed the brief?

Yes. In Wadsworth the two lawyers who signed without reviewing were fined $1,000 each. In Couvrette local counsel who did not sign the briefs was ordered to pay 15% of the fee award for failing to meaningfully participate, under the court's local rules. Rule 11 attaches to the signature.

How many cases are in Charlotin's database?

On 12 October 2026 the page said it was last updated on 10 October 2026 and counted 2,184 cases worldwide and 1,492 in the United States. It includes warnings as well as sanctions, so it is not a count of sanctions. Check the live page for the current figure.

Does ChatGPT make up legal cases?

It can. A language model predicts plausible text rather than looking up records. Stanford researchers found hallucination rates between 58% (ChatGPT 4) and 88% (Llama 2) on specific verifiable questions about random federal cases, and tools built for lawyers still hallucinated between 17% and 33% of the time in 2024 testing.

How do I verify AI legal citations before filing?

Open each case in a primary source, confirm that each quotation appears there word for word, read the case to confirm it supports your point, run a citator, and check statutes against the current official text. Then keep a record of what you checked. The steps above walk through it.

Is a legal research tool built for lawyers safe to trust without checking?

Not without checking. In Farris counsel used Westlaw's CoCounsel and the Sixth Circuit said the risk remains "even when new tools are sponsored by trusted legal technology providers." The duty to verify stays with the lawyer.

For the US statutes, regulations and court rules these laws sit in, see the US primary law API.

Arshita Anand

Co-Founder & CEO · Attorney

Arshita leads product and strategy at Vaquill AI, the US primary law data API. As an attorney, she decides which law we carry, how each section is sourced and cited, and what legal AI and compliance teams need from the data.