59 AI Hallucination Cases in Court Statistics for 2026

In a firm that litigates, an associate or a paralegal drafts the brief, sometimes with an AI tool open beside them, and a partner signs it.

When a citation in a signed brief fails and a judge notices, the court says so in writing. A case that does not exist, a quotation that is not in the opinion, a holding stretched past what the court decided: each one becomes a decision with the lawyer's name in it, and often the firm's.

We counted every decision of that kind in the public database that tracks them, as it stood on 9 November 2026, and sorted it by who filed, what went wrong, and what the court did.

The invented case is the part a firm already knows to look for. In most of the decisions about lawyers, the court found real authority misread or misquoted, and in 148 of them every authority cited was real.¹

The database holds only what a judge caught and wrote about, so it counts discovered errors, never how often they happen.²

Courts Have Written 2,286 Decisions on AI-Made Errors in Filings Since 2023

The count comes from the AI Hallucination Cases database kept by Damien Charlotin, a French lawyer and data scientist whose work a federal appeals court has cited in its own opinion.⁴ It records decisions in which a court or tribunal addressed AI-generated material beyond a passing reference and found, or implied, that a party relied on it.¹

By the date of the decision: 16 in 2023, 61 in 2024, 865 in 2025, and 1,344 in 2026 to 9 November 2026, with 0 undated¹. The newest months are never final, because decisions keep arriving after their date.

1,550 of the decisions come from US courts (67.80%), then Canada: 245, Australia: 112, and the UK: 91, across 46 jurisdictions.¹ Part of that share is American court records, which are easier to search than most; the maintainer says so himself.²

So compare whole years: 2026 has already passed all of 2025 for self-represented litigants, lawyers and judges alike.¹

917 of the 2,286 Decisions Involve a Lawyer

907 decisions name a lawyer among those responsible, and a handful more a prosecutor, a government lawyer, a paralegal or a federal defender: 917 decisions with a legal professional in all.¹ 1,314 name a self-represented litigant, 34 a judge, and 18 an expert witness.¹ A decision can name more than one.

None of these totals says which group errs more often. Nobody counts how many filings each group makes, so the record holds totals, not rates.

In a Stanford Law School study of US decisions up to October 2025, half of the firms behind them were solo practices, lawyers working on their own⁵.

Size did not keep larger firms out. K&L Gates, with more than 1,000 lawyers, is in that sample,⁵ over the brief in Lacey v. State Farm.

109 of 163 Dollar Amounts Against Lawyers Are $5,000 or Less

268 of the 917 lawyer decisions carry a monetary sanction (29.23%), against 123 of the 1,314 decisions about self-represented litigants (9.36%). 164 carry a professional sanction, such as a referral to the state bar, the body that licenses and disciplines lawyers (17.88%).¹

Where the database records a dollar amount against a lawyer, 163 times (the other monetary sanctions are in other currencies or of unknown amount), the median is $3,000 and 109 of the 163 are $5,000 or less; 13 are above $25,000.¹ The figure mixes fines with the other side's fees, and the three largest amounts are all mainly fee or costs awards.¹

The largest is Couvrette v. Wisnovsky in Oregon, $110,204.38: a $15,500 fine on the lead lawyer⁶ and $94,704.38 of the other side's fees³, after three briefs cited fifteen cases that do not exist, and the court dismissed the clients' claims for good. Local counsel, who had skimmed the first brief and read none of the later ones, still paid 15% of those fees, $14,205.66³.

The database's maintainer reads the sanctioned decisions this way: people are rarely sanctioned for the AI error alone, and usually for refusing to own up, doubling down, or blaming the intern.² The Fifth Circuit's $2,500 order of February 2026 named both: counsel failed to verify the brief, then was not forthcoming when asked.⁴

The Invented Case Is Only Part of What Courts Find in Lawyers' Filings

In the 917 lawyer decisions, courts found fabricated material in 736 (80.26%): a case, a statute or another source that does not exist. They found real authority misrepresented in 426 (46.46%), a case cited for a point it does not make, and false quotations in 301 (32.82%), words attributed to a real court that are not in its opinion.¹

450 of the decisions carry more than one kind, so the shares overlap and never add up to the whole.¹

Count the two kinds that involve real authority together and they reach 550 of the 917 lawyer decisions, 59.98% of them.¹ Case law is the material in nearly every decision of the database, 2,047 of the 2,286, ahead of statutes and rules at 262 and exhibits at 166.¹

A Check That Only Asks Whether a Case Exists Misses 148 Lawyer Decisions

In 148 of the 917, 16.14%, every authority the filing cited was real. The court still found it misread, or found words attributed to it that it never wrote.¹

That share was 39 of 361 lawyer decisions in 2025 (10.80%) and 106 of 525 in 2026 (20.19%).¹ The database cannot say why it rose: more careful filers and more careful classification of new entries would both produce it.

A tool or a paralegal that only confirms each cited case exists finds nothing to flag in these filings, because every case does. Opening the case at the cited page and reading the passage against the sentence that relies on it, checking each citation and quotation against the opinion it points to, is the check that can catch them; whether a real holding stretches as far as the brief needs is still a lawyer's judgment.

In Lacey v. State Farm, lawyers at two firms built a brief from an AI-assisted outline that no one checked before filing; about nine of its 27 citations were wrong in some way, and the two firms were ordered to pay $31,100 between them⁷.

What a firm pays for an AI error follows from whether anyone read the cited passages before the partner signed. In Couvrette and in Lacey, nobody did.