AI citation errors: a verification protocol for legal teams

Damien Charlotin keeps an independent public database of decisions in which courts or tribunals have addressed generative-AI hallucinations or, in some entries, alleged AI use. The page was last updated on 5 September 2026 and listed 2,022 legal cases when accessed on 9 September 2026. It remains a work in progress and it is not an official court registry, so the count should not be read as a complete judicial total. The scale is enough for the point I care about, which is an operational one: AI-assisted legal work needs a verification record before it leaves the team.

One decision states the point plainly. In LNU v. Blanche, decided on 3 June 2026, the US Court of Appeals for the Ninth Circuit suspended two attorneys from practice before that court for six months, imposed a $2,500 sanction on each, and required AI-use disclosure and certification in their filings for two years. The sanction is the headline, but the reasoning is what anyone building a process should read twice: the relevant duties are breached when an unchecked document is signed and filed, not merely when AI is used for research or drafting. Türkiye has its own example: in a Kısa Dalga interview of 17 June 2026, a lawyer described a petition filed before a first-instance court that carried Yargıtay citations with fabricated case numbers and said the court wrote to the prosecutor and the bar the same day; those are the interviewee’s statements, not a court finding.

That lesson is narrower than a general warning against AI, and more useful, because a warning doesn’t survive contact with a deadline. What a team has to be able to show is which source was opened, which quotation was checked against it, which proposition was tested, and who accepted responsibility for the final text.

The protocol I use has five steps, and none of them is optional for anything that leaves the team. One, open every citation in the official reporter or database; never ask a second AI whether the first AI’s case exists. Two, compare every quotation against the opinion word by word. Three, test the proposition, because the case must actually say what the draft claims it says. Four, test validity, because the authority must still be good law. Five, a named person signs the verification note before the document goes out. Court filings are not the only place this bites; client memos and advice go through the same five steps. It is a dull list, and I haven’t yet found the step I would be willing to drop.

In-house teams have one extra lever, and it is the engagement letter with outside counsel. Set out there how the firm verifies AI-assisted citations and who certifies the review, and nobody has to raise the question awkwardly after a document has gone out. The Union of Turkish Bar Associations pointed the same way when it published its advisory guide on AI use by lawyers on 28 August 2026: human oversight, verification of outputs, professional secrecy, and the lawyer’s final responsibility. A tool can carry a good deal of the workflow, but the signature at the bottom is still a person’s, and so is the judgment behind it.

Next: Marketplace intermediary liability in Turkey: three regimes to read together →