Since 2023, courts across the United States (and an alarming number of places that should know better) have been processing a new category of motion: the one where opposing counsel explains that six cases in your brief do not exist, never existed, and appear to have been dreamed up by a language model with a gift for Bluebook cosplay.
Researchers now track on the order of 1,400 court and tribunal decisions worldwide confronting AI hallucinations in filings — about 955 in U.S. courts alone as of mid-2026. Sanctions have escalated from "please don't do that again" to suspensions, disqualifications, bar referrals, and the judicial equivalent of being sent to your room without appellate relief.
This is the funny version of that timeline. The cases are real. The citations in the briefs were not.
Origin story: Mata v. Avianca (the serving cart heard round the world)
June 2023. Southern District of New York. Roberto Mata sued Avianca because a serving cart allegedly injured his knee — a completely ordinary personal injury case until his lawyers, Steven Schwartz and Peter LoDuca of Levidow, Levidow & Oberman, filed an opposition brief citing six prior airline-liability cases as precedent.
None of the six existed. The fabricated roster included names like Varghese v. China South Airlines, Shaboon v. EgyptAir, Petersen v. Iran Air, and Estate of Durden v. KLM Royal Dutch Airlines — each complete with fake docket numbers, invented quotes, and legal reasoning that sounded like a judge wrote it after a long flight and a shorter fact-check.
When the court started asking questions, Schwartz admitted he had used ChatGPT. He had even asked ChatGPT to confirm the cases were real. ChatGPT replied that they "indeed exist" and could be found in reputable legal databases such as LexisNexis and Westlaw — which is the AI equivalent of "trust me bro" with footnotes.
Judge P. Kevin Castel called them "bogus judicial decisions with bogus quotes and bogus internal citations," imposed $5,000 in sanctions on both attorneys, and dismissed the underlying injury case. Mata is now the citation every other judge uses when they're about to ruin someone's afternoon.
Federal courts: the montage gets expensive
Ellis George LLP & K&L Gates (C.D. Cal., May 2025)
Two major firms submitted a supplemental brief to a Special Master packed with hallucinated citations. The Special Master struck the offending brief, denied the relief they sought, and ordered $31,100 in opposing fees. He also noted he "almost cited fake material in a judicial order" while reviewing their work — which is how you know you've achieved something special.
Johnson v. Dunn (N.D. Ala., July 2025)
Butler Snow LLP — a large firm with AI policies and an actual AI Committee — watched three attorneys file motions containing ChatGPT citations nobody verified. Judge Anna Manasco didn't bother with a fine. She disqualified all three attorneys from the case and referred them to bar regulators. Message received: your compliance deck is not a substitute for opening Westlaw.
United States v. Ramirez (N.D. Ind., 2025)
After an "exhaustive review," the court found fabricated cases not just in the brief everyone was yelling about, but in two additional briefs. Recommended sanctions: $15,000. The court noted that AI-generated excerpts looking valid to Ramirez "does not relieve him of his duty to conduct a reasonable inquiry" — which is judge for "it looked real to me" is not a defense to a motion to dismiss.
Hoverboard / Walmart (D. Wyo., 2025)
Faulty lithium battery. Eight non-existent cases. One attorney lost his pro hac vice admission; three others paid $1,000–$3,000 each. Wyoming: come for the scenery, leave your imaginary precedent at home.
Wage & hour class action (N.D. Cal., 2026)
AI-generated citations in a settlement approval motion. Sanctions: $250 to the clerk, motion stricken without leave to refile, referral to the Standing Committee on Professional Conduct. The court emphasized that good intentions don't save you from Rule 11. The robot doesn't sign the certificate. You do.
State courts: when the appellate record is fan fiction
Alabama Supreme Court (April 2026)
The court dismissed an entire appeal as frivolous after briefs "riddled with fabricated legal citations" from an AI tool. Sanctions included $17,200 in fees to the appellee, double costs, a requirement that future filings be co-signed by another attorney, and a bar referral. Best detail: after being told he'd cited a made-up case, the lawyer promised it wouldn't happen again — then cited nonexistent cases "at the end of the very next sentence." That's not a learning curve. That's a learning flatline.
State v. Payne — Georgia Supreme Court (May 2026)
This one stopped being funny halfway through. A prosecutor's AI-generated fake citations made it into a brief on a murder case. The trial judge adopted much of her proposed order — fake cites included — in denying a new trial. The Georgia Supreme Court admonished the DA's office, suspended the prosecutor for six months, vacated the trial court's order, and ordered ethics and AI CLE. When hallucinated law reaches criminal outcomes, the punchline is someone's liberty.
Mostafavi v. Land of the Free (Cal. Ct. App., Sept. 2025)
Twenty-one of twenty-three quoted passages in an appellate brief were fabricated. Fine: $10,000 — described at the time as the largest AI-related sanction from a California state court. The panel noted the attorney wasted the court's time and taxpayers' money. Accurate, if understated.
Ninth Circuit, June 2026: the national standard arrives
The largest federal appellate court in the country issued its first major AI-accountability ruling on June 3, 2026. Attorneys Mike Sethi and William Rounds filed briefs with "multiple nonexistent cases, misattributed quotations, and gross misrepresentations of real cases." Sethi amended a filing to remove fake cases without disclosing they'd been fake — which the court called "subtle deception," a phrase that should keep litigation support up at night.
Sanctions: $2,500 each, six-month suspension from practicing before the Ninth Circuit, and a two-year requirement to disclose any generative AI used in filings. The court sent a disciplinary notice to the entire bar: using AI isn't the sin. Failing to verify it — and failing to be candid when caught — is.
Bar discipline and the Phoenix Suns (Arizona edition)
Courts aren't the only ones keeping score. In April 2026 the California State Bar filed charges against three attorneys for AI-generated fake citations — with recommended probation, suspensions, and possible disbarment pending. The Bar's message: the duty of competence doesn't transfer to the tool.
Closer to home: a federal judge in Arizona disciplined attorneys in a Phoenix Suns discrimination lawsuit, finding 18 instances of incorrect or fabricated case references across three filings — "likely generated or researched using AI." Attorney Sheree Wright was ordered to pay fees that could reach hundreds of thousands of dollars, and both attorneys were ordered to attend training on ethical AI use. Even the Suns organization wasn't the only thing getting roasted in Phoenix that week.
Patterns (or: how to lose a bar license in one easy step)
Every case shares the same plot:
- Non-verification. Nobody checked Lexis, Westlaw, or a free docket.
- Plausible fiction. Realistic party names, believable quotes, perfect vibes.
- Compounding. Deny, double down, or file more fake cites after being warned.
- Good faith doesn't save you. Rule 11 cares about inquiry, not your feelings.
- Big firms aren't immune. AI committees don't verify briefs while you sleep.
Statistically, this is still rare — roughly one documented hallucination case per 41,000 U.S. filings since 2023. But "rare" is doing a lot of work when you're the one explaining to a judge why Shaboon v. EgyptAir has no docket number. Pro se litigants account for a disproportionate share; among attorneys, the court discovers the problem more than 60% of the time — not opposing counsel. When opposing counsel catches it, you might get a meet-and-confer. When the judge catches it, you're already in the part of the docket where they discuss your bar number.
The sanctions escalation chart (abridged)
- 2023: Famous fines, stern lectures, Mata enters the chat.
- 2024: More money, state suspensions, appellate referrals.
- 2025: Disqualifications, $15K+ hits, pro hac vice revocations.
- 2026: Whole appeals dismissed, six-month court bans, murder-case do-overs, bar charges.
Courts aren't banning AI. They're banning unverified output and uncandid cleanup. Several now require disclosure of AI use. The Ninth Circuit's two-year disclosure order may be a preview of what's coming everywhere.
The workflow that keeps you off this list
Retrieve first. Cite second. Show the queries. Quote verbatim from text you actually pulled. Document what you couldn't find. Run a verification pass that flags anything not in your session corpus.
That's why Lexora exists — not because AI can't help with legal research, but because the help has to be grounded, traceable, and honest about gaps. If your tool can't show you the retrieval path, you're not researching. You're speed-running a sanctions order with better typography.
The robots will write you a beautiful brief. The courts have read the executive summary. Pick a side.
Try Lexora · The hallucinated citation problem · Why retrieval-first