A running record of real, published AI accuracy failures across law, journalism, and research. Every item links to the primary coverage — we don't ask you to take our word for it. We read each one at the source before posting it.
Mata v. Avianca. Judge P. Kevin Castel sanctioned two attorneys and their firm after a brief cited six cases that did not exist. The detail that still matters most: when challenged, counsel asked ChatGPT whether the citations were real, and it said yes. The court found bad faith not in the original error but in defending it afterwards — the pattern every case since has repeated. Case summary.
Sullivan & Cromwell filed documents in the Prince Group wind-down before Judge Martin Glenn (S.D.N.Y.) containing wrong and fabricated citations and inaccurate article titles. The firm apologised ten days later. It was opposing counsel — Boies Schiller Flexner — who caught it, not the firm's own review. Seniority is not a control. Global Legal Post · Canadian Lawyer.
A report for Australia's Department of Employment and Workplace Relations, on a AUD 440,000 contract, contained a fabricated quote from a Federal Court judgment and references to academic papers that do not exist. Deloitte issued a partial refund and the revised version disclosed that a generative AI system had been used. A university researcher, not the client and not the firm, found it.
A researcher tracking these filings watched the rate go from roughly two a week to two or three a day through 2025. The fabricated citations carry real case numbers and the initials of real judges — the judges exist; the cases don't. That's the mis-attribution failure an existence check can't catch.
An attorney used ChatGPT to "enhance" an appellate brief and didn't verify the citations. The court found nearly every quotation invented, imposed the fine, and referred him to the state bar. The tooling to catch this before filing is a citation check against the register of record.
Verification is no longer just about your own draft. Courts are now faulting counsel for not detecting fabricated citations in the other side's filing. When checking every citation — yours and theirs — becomes a professional duty, doing it by hand stops being viable.
A freelancer used AI to research a syndicated summer reading list; 10 of the 15 titles were invented — including a "climate fiction novel" attributed to Isabel Allende that she never wrote. It ran in the Chicago Sun-Times and the Philadelphia Inquirer. Every fabricated title fails a single check against the book registry.
A 53% correction rate, with at least five substantive errors in a single compounding-interest explainer. The outlet paused the tool and added disclosure. The lesson for any desk: AI-assisted drafting needs an accuracy layer before publication, not a correction after.
Bylines and profile photos were traced to an AI-portrait marketplace; the outlet deleted the articles. A reminder that the reputational cost of unverified AI content isn't a bad day — it's the story becoming about you.
The one preregistered study of the commercial tools, tested over their own case-law databases — the exact setup vendors say fixes hallucination. It didn't. Retrieval-augmentation reduced the rate; it did not remove it.
Every item above was read at the linked source before posting. We link to primary coverage, not aggregators, wherever possible. If a link breaks, tell us — a dead link on a page about accuracy is not one we'll leave up.
AttestAlly — because “the AI cited it” isn’t a defence.