AI hallucinations in legal work: sanctions, cases, and how to avoid them

A researcher named Damien Charlotin has spent since April 2025 doing something most law firms should be paying him to do: cataloguing every court filing anywhere in the world where a lawyer got caught submitting AI-fabricated legal citations. His public database now sits above 1,700 documented incidents. Most of the coverage of that number treats it as a curiosity. It isn't. It's a map of exactly where courts draw the line, and the line has moved twice in 2026 already.

Here's what most write-ups on this topic get wrong. They treat every hallucination case as the same story: careless lawyer, fake case, judge gets angry, fine gets paid, everyone moves on. The actual sanctions orders from this year show three different things courts are punishing, and only one of them is the hallucination itself.

The scale, in one number

U.S. courts imposed more than $145,000 in sanctions in the first quarter of 2026 alone for AI-generated fabrications in court filings. That figure covers one quarter, in one country. Charlotin's tracker counts incidents going back to early 2025 and shows the pace accelerating, not slowing, as more lawyers adopt AI drafting tools without adopting the verification habits to match. Pro se litigants make up the largest single group in the database, but attorneys, the people with a bar license and a professional duty of candor, account for a share large enough that no state bar can treat this as a fringe problem anymore.

Nebraska: the case that ended a career

In February 2026, Omaha attorney Greg Lake filed a brief in a divorce appeal. Of the 63 citations in it, 57 were defective. Twenty cited cases that don't exist at all. The Nebraska Supreme Court suspended him indefinitely in April and ordered a full disciplinary investigation. This wasn't a brief with one bad citation slipping through review. This was a brief where review apparently never happened, and the ratio is what turned a sanctions matter into a career-ending one.

The Ninth Circuit changed the rule for everyone

The more important case this year is quieter and stranger. In June 2026, a Ninth Circuit panel sanctioned two Orange County immigration attorneys, Mike Sethi and William Rounds, over briefs containing citations to opinions that don't exist and quotes fabricated from real cases. The published order fined each attorney $2,500 and suspended both from practicing before the court for six months. The reasoning matters more than the fine. The panel said explicitly that it was not sanctioning the attorneys for using generative AI to research or draft. The violation happens at the point of signing and filing something you didn't personally verify, not at the point of typing a prompt.

The order goes further than any prior ruling on this issue. It requires every future filing from that firm to include a sworn statement disclosing whether generative AI was used, naming the tool, and certifying under penalty of perjury that the signing attorney personally checked every citation and quotation against real authority. That requirement applies to one firm today. Expect other circuits to start requiring something close to it within the next year, the same way electronic filing certifications became routine after courts got tired of asking nicely.

Mississippi: when both sides get caught

A federal judge in the Northern District of Mississippi handled a contract dispute where the plaintiff's counsel and the city's counsel had each used AI to draft filings and each failed to catch the hallucinations before submission. Both lawyers lost their permission to appear before that court for two years and were fined $2,500 and $3,500. Neither side could point to the other and claim the problem was isolated. That should worry any lawyer who assumes opposing counsel's sloppiness is somehow their protection.

Oregon set a different kind of precedent in December 2025 with a per-infraction fee schedule: $500 for each fabricated citation, and $1,000 for a fabricated quotation, applied item by item rather than as one lump penalty. Federal courts in the same jurisdiction have since adopted the same formula. A brief with five fake citations and two invented quotes under that schedule runs $4,500 before anyone even gets to the merits of the underlying case.

What actually stops this

None of the lawyers in these cases set out to file fake law. They trusted a tool that sounds confident whether it's right or wrong, and skipped the one step that would have caught it. The fix isn't avoiding AI. It's treating its output the way you'd treat a brief from a first-year associate you've never worked with before: useful, possibly excellent, and unverified until you've read the source yourself.

1. Open every case an AI tool cites and read the actual holding, not just the case name and reporter citation.
2. Treat quotations as unverified until you've found the exact language in the source document, not a paraphrase that sounds right.
3. Keep a record of which filings used AI assistance and who reviewed them, so you can show a verification process existed if a citation ever gets questioned.
4. If you find an error after filing, correct it and disclose it before anyone else does. Every case above where the lawyer self-reported got a lighter sanction than the ones where the court found the problem first.

This is the same candor duty covered under Model Rule 3.3, which our breakdown of ABA Opinion 512 goes into in more detail, including the four other duties that apply alongside it. If your firm doesn't have a written policy covering AI use in drafting and research, that's the document these verification habits belong in, and our AI policy template gets a usable draft in front of you in an afternoon rather than a month.

Where this is heading

The Ninth Circuit's disclosure requirement is the signal to watch. Once one circuit starts requiring a sworn AI-use certification on every filing, plaintiffs' bars and defense bars alike start drafting model language for it, and judges in other circuits start asking why they aren't requiring the same thing. Expect a version of this certification to become close to standard practice within two years, the same way courts standardized electronic signature blocks once a few districts started requiring them. Get ahead of it now and it costs you a paragraph in your engagement letter. Get caught behind it and it costs you six months of practice rights and a fine that scales with every fake citation in the brief.

Get a ready-to-use AI-use disclosure paragraph and a verification checklist mapped to these cases in The Lawyer's AI Starter Kit, free. And if you're choosing which AI research tool is actually built to minimize this risk, the best AI tools for lawyers page tracks which ones ground answers in real case law and which ones don't.

FAQ