2,225 cases and counting: the AI hallucination problem is not slowing down

When Mata v. Avianca hit in 2023, most of us treated it as a story about two lawyers who got unlucky. It was the first of its kind, it made the news, and the assumption was that the profession would learn from it.

The profession did not learn from it fast enough.

The count

Damien Charlotin, a researcher in France, keeps a public database of court decisions where AI-generated fabrications showed up in a filing or a ruling. As of October 27, 2026, it lists 2,225 matters. Roughly 1,283 involve self-represented litigants. About 879 involve lawyers. Judges account for 32. Experts account for 17.

Two things about that number.

First, it is a floor. The database only counts matters where a court noticed and wrote it down. A fake citation that opposing counsel quietly caught and the judge never addressed does not appear.

Second, the lawyer count is the one that should bother you. Self-represented litigants are not licensed and not insured. Lawyers are both.

The sanctions are not a slap on the wrist anymore

In 2023, Mata produced a $5,000 sanction. That looked serious at the time. Look at what has happened since.

• Gauthier v. Goodyear (E.D. Tex., Nov. 2024): $2,000 and a required CLE course. Light, but the court made clear it would not stay light.
• Wadsworth v. Walmart (D. Wyo., Feb. 2025): eight fake citations, fines of $3,000, $1,000 and $1,000, and the drafting lawyer's pro hac vice admission revoked.
• Lacey v. State Farm (C.D. Cal., special master, May 2025): $31,100 against two firms jointly, after roughly 9 of 27 citations turned out to be wrong.
• Noland v. Land of the Free (Cal. Ct. App., Sept. 2025): $10,000, a published opinion, and a State Bar referral.
• Couvrette v. Wisnovsky (D. Or.): about $110,000 in total and dismissal with prejudice, after an attempted cover-up.
• Twigg v. BSN Sports (M.D. Pa., about June 2026): a six-month suspension.

The pattern is simple. Courts started with fines and CLE. They moved to fee-shifting, disqualification, referrals, and now suspensions. In the Fifth Circuit, the court in Fletcher v. Experian (Feb. 2026) said what a lot of judges are thinking: ignorance of the risk is no longer an excuse.

Why it keeps happening

Most of these lawyers do not appear to be lazy or dishonest. The tool they used is built to produce text that sounds right, and the lawyer's process had no step that catches text that sounds right but is not.

A general-purpose model has no connection to a case database. It does not know whether a citation is real. It has no way to tell you which sentence it is guessing at. And the review step depends entirely on a tired lawyer remembering to pull every cite at 11:00 at night.

Even a firm that does everything "right" can get hit. In Johnson v. Dunn (N.D. Ala., July 2025), the firm had a written AI policy and had warned its lawyers. A partner's name was still on the signature block, and the court sanctioned him, disqualified the lawyers and referred them to the bar. A policy that lives in a handbook is not a control.

What a control actually looks like

A control is something that sits in the path of the work and stops the bad output. Not a reminder. Not a training video.

This is the reason JILL is built the way it is.

Citations are validated before they reach you. Every citation JILL produces is checked against authoritative case databases for the matter's jurisdiction. A case that does not exist does not appear in the output. This is a gate, not a suggestion to verify later.

Every paragraph carries a confidence score. You can see which parts of a draft rest on strong, current authority and which are thin. That tells you where to spend your review time instead of spreading it evenly across fifty pages.

Everything is logged. Each prompt, retrieval, and output is recorded in an append-only, signed audit log under the Discover Docket Ethical and Accountability Standards. If a court or a bar counsel asks what you did, you can show them.

None of that replaces your judgment. You still read the draft. You still sign it. What changes is that the failure that put 2,225 matters in a database is blocked before it gets to your signature.

The bottom line

The question for your firm is no longer whether AI hallucinations are real. A court has probably addressed one this week. The question is whether the way your lawyers use AI today would survive a judge asking, "Show me how you checked this."

If the honest answer is "someone was supposed to check," that is the gap.

See how JILL validates authority →

Built for ABA Formal Opinion 512 →