AI Hallucination Sanctions Timeline: Every Major Case from Mata to 2026

On 22 June 2023 Judge P. Kevin Castel fined two lawyers and their firm $5,000, jointly and severally, for citing six cases ChatGPT had invented. Mata v. Avianca was modest by design. Two years later Judge Anna Manasco wrote that “if fines and public embarrassment were effective deterrents, there would not be so many cases to cite”, and in March 2026 the Sixth Circuit added that “smaller fines have plainly been inadequate” before imposing $15,000 on each of two lawyers plus their opponents’ fees and double costs.

That is the shape of AI hallucination sanctions cases between 2023 and 2026: a curve of decisions that rose from a handful a year to more than a hundred a month, and a ladder of consequences that climbed from fines to referrals, suspensions, dismissed claims and one cancelled trial. Every rung has a name, a date and a quotable line, and they are all below.

If you are here because you have just found a fake citation in something you filed, stop reading this and go to what to do in the next 48 hours. Everyone else: this is the page to open when a partner says “that only happens to solos with ChatGPT”.

The count: 2,039 decisions and the March 2026 peak

Damien Charlotin’s AI Hallucination Cases database is the reference. It lists only decisions in which a court explicitly found, or implied, that a party relied on hallucinated material, so allegations that were never ruled on are excluded and the true number is higher. On 12 September 2026 it stood at 2,039 decisions: USA 1,396, Canada 217, Australia 110, UK 69, Israel 57, Brazil 41, Italy 15, India 15, France 13, Germany 11, Austria 2.

The year-by-year totals from the database’s own CSV are the statistic to remember: 16 decisions in 2023, 61 in 2024, 851 in 2025 and 1,111 in 2026 up to 12 September.

Two things stand out. The peak, 179 decisions in March 2026, came after the frontier models had improved dramatically on every legal benchmark, which is the paradox explored in why AI makes up fake cases. And the lawyer-involved share has held at roughly 40% throughout; this is not, as is often claimed, purely a self-represented-litigant problem. On the nature of the errors, the database counts 1,689 fabricated authorities, 853 misrepresented ones, 549 false quotations from real cases and 33 instances of outdated advice.

Charlotin himself told an interviewer in September 2026 that the curve had “reached a plateau”, attributing it to “better tools and higher awareness” (as reported by a syndicated Daily Caller News Foundation piece). A plateau above 100 decisions a month is not comfort. Eugene Volokh counted 17 US decisions noting suspected hallucinations on a single day, 31 March 2026.

2023: Mata v. Avianca and the first suspension

The story everyone knows is worth retelling precisely, because the details are the lesson. Steven Schwartz’s firm had no Westlaw or Lexis subscription, only a limited Fastcase plan. He asked ChatGPT for authority against Avianca and it produced Varghese v. China Southern Airlines and five others. When the court demanded the opinions, he asked ChatGPT for copies and filed those too. His own exhibits show him asking “Is Varghese a real case” and being told it “can be found in reputable legal databases such as LexisNexis and Westlaw”. At the hearing Judge Castel read a fabricated opinion aloud and asked, “Can we agree that’s legal gibberish?”

Judge Starr’s order, entered a week after the Mata affidavit surfaced, was the first in the US. It requires a certificate that no part of a filing was drafted by generative AI, or that any AI-drafted language “was checked for accuracy, using print reporters or traditional legal databases, by a human being”. By April 2026 one tracker counted 113 such orders in force; the court standing orders guide explains how to check your judge.

2024 to early 2025: Park v. Kim, Kohls and the first firm-level fallout

2024 was quiet on the docket, 61 decisions in the whole year, and loud in the appellate courts. The Second Circuit referred a lawyer to its grievance panel in January; the Fifth Circuit in June declined to adopt an AI certification rule because the existing rules already covered it, adding that “‘I used AI’ will not be an excuse for an otherwise sanctionable offense”. Then, in the first two months of 2025, two cases changed who the problem belonged to.

Wadsworth matters for a different reason. The eight fake cases did not come from ChatGPT; the court expressly found “It appears Plaintiffs’ attorneys did not use ChatGPT”. They came from MX2.law, the firm’s own platform, after the drafting lawyer typed “add to this Motion in Limine Federal Case law from Wyoming setting forth requirements for motions in limine”. Two partners e-signed without reading. The firm then emailed more than 1,000 of its lawyers that fabricated citations could mean termination.

2025: Lacey, Johnson v. Dunn, Coomer, Noland, Ayinde

2025 is the year the numbers exploded (851 decisions) and the year the “small firm, consumer chatbot” explanation died. The fourteenth-largest US firm, a 400-plus-lawyer Alabama-based firm with a written AI policy, a Latham associate, the MyPillow legal team, a pupil barrister in London and a family-law brief in Cologne all appear.

The two cases to put in front of any partner who still thinks a policy is a control are Johnson v. Dunn and Lacey. Butler Snow had a written AI policy since June 2023 and an AI committee. Judge Manasco: “They benefitted from repeated warnings, internal controls, and firm policies about the dangers of AI misuse… And yet here we are.” She spared the firm, which had Morgan Lewis check more than 2,400 citations in 330 filings (none wrong), and disqualified the three individuals. In Lacey, the AI-assisted outline came from co-counsel at Ellis George; K&L Gates folded it into the brief unread, and a revised brief still contained errors.

Noland added what LawSites called a “new wrinkle”: the respondents won the appeal and were denied their fees because they “did not alert the court to the fabricated citations”. Cite-checking your opponent is now part of the job, which is why there is a separate guide on responding to AI-generated filings from the other side. The European decisions in the table get their own treatment in AI hallucination cases in the UK, Germany and Europe.

2026: Couvrette, Whiting, Prososki, suspensions and referrals

By 2026 the appellate courts had run out of patience with fines. The Sixth Circuit said so in terms; the Ninth Circuit suspended; a Nebraska appeal was dismissed outright; and in Mississippi a trial was cancelled because both sides had filed fakes.

Note what Withers is not: the plaintiff’s surname, not the London law firm, and the two-year measure is a bar on appearing in that district after pro hac vice revocation, not a bar suspension. Note also what Prososki is not: the March opinion referred Mr Lake to the Counsel for Discipline; the April suspension rests on Nebraska press reports, and the “57 of 63 citations” figure circulating online is opposing counsel’s count, not a court finding.

The quietest entry of the year may be the most instructive. In April 2026 Sullivan & Cromwell, which advises OpenAI, sent a three-page single-spaced letter to Chief Judge Glenn of the S.D.N.Y. bankruptcy court listing “inaccurate citations and other errors” in a motion, spotted by Boies Schiller. No sanction was reported. David Lat’s verdict: “There but for the grace of God go (A)I.”

Who gets caught in AI hallucination sanctions cases: firm size and party type

Stanford’s Riana Pfefferkorn analysed 114 US lawyer cases from the database in October 2025. Of the 129 firms and entities involved, solos were 50.4%, firms of 2 to 25 lawyers 39.5% and firms of more than 1,000 lawyers 1.6% (K&L Gates and Morgan & Morgan). Plaintiff-side lawyers accounted for 56% of cases and defence 31%. Of the 34 cases that named a tool, 18 involved ChatGPT, followed by Westlaw AI, Claude, Copilot, Gemini and Lexis (Stanford CIS).

Read those numbers with the base rates in mind. A 50% share is not proof that solos are careless; it is proof that a firm without a research subscription and without a second reader is where an unverified draft goes straight to the court. The practice areas that recur are the high-volume, thin-margin ones: immigration (Lnu, Dehghani), consumer bankruptcy (In re Martin), personal injury (Wadsworth), family law (AG Köln, KG Berlin, Prososki). BigLaw fails differently: experts, formatting requests and unread co-counsel drafts.

The 32 judge entries deserve a line. Two federal district judges told Senator Grassley in October 2025 that a clerk had used Perplexity and an intern ChatGPT to draft orders that were docketed with fake parties and misquotes; Judge Wingate’s remedy is one any firm can copy: a second reviewer, and every cited case printed from Westlaw and attached to the draft.

The escalation ladder: from admonition to a cancelled trial

Put the cases in order of severity rather than date and a ladder appears. Courts start at the bottom and move up when the conduct after discovery is worse than the conduct before it.

What judges quote back

The opinions are now quoting each other. The sentence most often repeated originated in California and was picked up by the Sixth Circuit in Whiting:

“Simply stated, no brief, pleading, motion, or any other paper filed in any court should contain any citations—whether provided by generative AI or any other source—that the attorney responsible for submitting the pleading has not personally read and verified.” — California Court of Appeal, Noland v. Land of the Free (12 September 2025)

On the standard of care:

“[A] competent and diligent attorney must do more than prompt generative AI, check that the citations provided by the AI are real and the subject matter roughly on point, and call it a day. … A competent and diligent attorney must also read and reason.” — Ninth Circuit, Lnu v. Blanche (3 June 2026)

On the excuse that the tool was a good one:

“no matter how much one pays for ‘premier’ or ‘corporate’ versions of AI products, it does not negate an attorney’s obligation to verify all citations of authority” — Illinois Appellate Court, Scott v. Illinois Human Rights Commission (28 July 2026)

On the excuse that the tool said so:

“Citing nonexistent case law or misrepresenting the holdings of a case is making a false statement to a court. It does not matter if [generative AI] told you so.” — Nebraska Supreme Court, Prososki v. Regan (20 March 2026)

And from the bench in Santa Fe, to a lawyer whose ChatGPT-assisted brief in a murder appeal quoted witnesses who did not exist:

“Counsel, do you watch the news? Do you listen to the radio? Do you read anything about what’s going on in the world? Because the problem with lawyers relying on AI hallucinations is an ‘above-the-fold’ story every single day.” — Justice Shannon Bacon, New Mexico Supreme Court (The Guardian, 11 September 2026)

The UK equivalent is Dame Victoria Sharp’s paragraph 6 in Ayinde: freely available tools “are not capable of conducting reliable legal research”, and lawyers who use them anyway “have a professional duty therefore to check the accuracy of such research by reference to authoritative sources”. The SRA’s warning notice of August 2026 reduces it to one sentence: “Reliance on an output of AI would not be a suitable defence.”

How to use this page in a CLE or firm training

A timeline is only useful if it changes what someone does on Thursday afternoon. Three ways to turn it into practice, each with a prompt you can paste.

Run the cases as exercises, not slides. Paul Weiss found its first PowerPoint session “ineffective” and moved to hands-on work; Texas lawyers asked their bar for training “similar to a lab you might have in biology class”. Hand participants the Wadsworth prompt (“add to this Motion in Limine Federal Case law from Wyoming”) and a research tool, and have them find which of the resulting cases exist. Then hand them the Latham request (“provide a properly formatted legal citation”) and watch a real article acquire a fake author. That is how the sessions in AI Lab for Lawyers open: participants find the error on their own screen before a judge does, and the habit sticks in a way a case list never will.

You are helping me design a 20-minute verification exercise for lawyers in my firm, based on a real sanctions decision I will paste below.
From the decision, extract: (1) the exact prompt or request the lawyer made, if the court records it; (2) the number and type of defective citations (fabricated, misrepresented, false quotation); (3) what the lawyer said when challenged; (4) the sanction and the court's stated reason for its severity.
Then draft: a one-paragraph scenario that puts a participant in the lawyer's position the day before filing; three questions that force them to decide what they would check and how; and a model answer keyed to the six verification layers (existence, quotation, holding, status, jurisdiction, documentation).
Do not invent any fact about the decision that is not in the text I paste. Do not generate any fake citations for the exercise; I will supply those myself.

Decision:
[paste]

Build the verification table before you file anything. The model is allowed to list; it is never allowed to verify. This is the pattern from the prompt library, and it is the same table a court will want to see if something goes wrong.

List every case, statute, rule, regulation and secondary source cited in the document below in a table with these columns: Citation as written | Proposition it is cited for (quote the sentence) | Pinpoint given? (Y/N) | Quotation present? (Y/N) | Red flags (reporter or volume mismatch, implausibly on-point case name, suspiciously perfect quotation, court or judge that may not exist).
Do not tell me whether any citation exists or is correct; I will check each one in a primary database. Do not add, correct or reformat any citation. If the document contains a bracket placeholder such as [cite], flag it in a separate list.

Document:
[paste]

Check the judge’s order before you check the citations. Judge Starr’s certificate, the Federal Court of Canada’s first-paragraph declaration and Lnu’s two-year sworn disclosure are three different regimes, and 113 active orders sit between them.

For a filing in [court, judge], using only the court's website at [URL] and the order text I paste, tell me: whether this judge or court has a generative-AI standing order, certification or disclosure requirement; what the certificate must say, quoted; whether the tool and the affected portions must be identified; and the consequence stated for non-compliance. Then draft the certificate in the required form for a filing in which AI assisted with [describe] and every citation was verified by [name] in [database] on [date].
If you find no order in the sources provided, write "NO ORDER FOUND ON THE SOURCES PROVIDED" and stop; do not infer one from other courts.

Order text:
[paste]

Keep the record. Underwriters have started asking about it. ALPS’s Chris Newbold said in August 2026 that carriers are “not seeing a lot of what I would call direct AI claims at this point”, but underwriters have started asking “Do you police it?”, and a pre-filing verification log is on the list of renewal artefacts; see malpractice insurance and AI mistakes. Lee Norcross of L Squared put the reason bluntly: “AI does not have a law license to lose.”

Where to go next: the six-layer verification protocol is the fix; what to do in the first 48 hours after finding a fake citation is the emergency plan; the European sanctions digest covers Ayinde, the UKUT and the German courts in depth; and the wider numbers on adoption and risk sit in legal AI statistics. The whole verification cluster is built around one sentence from the Ninth Circuit: read and reason. If you would rather build that habit live, on your own documents, with the tools open, that is what the four sessions of AI Lab for Lawyers are for.