The AI Lied About You. In Court, That Might Not Matter.

2026-07-24

In May 2025, a Georgia court threw out one of the first defamation lawsuits brought against an AI company. The claim itself was not in dispute: ChatGPT had told a user that radio host Mark Walters was accused of embezzling funds from a gun-rights organisation. It was false, and easy to show as false. Walters lost anyway.

He lost because the only person who ever saw the statement was a journalist testing the system, who recognised it as untrue on sight and never published it. With no second reader, no spread, and no evidence that anyone's opinion of Walters had changed, the court found there was nothing left to compensate. Two more cases making similar claims, against Google and against xAI, are still working through the courts. Neither has produced a plaintiff win. Neither has produced a clear rule other than the one Walters already learned the hard way.

That pattern is the real story, and it's a more useful one than "AI defamation is unwinnable." Defamation law was built for a world of publishers: newspapers with circulation figures, broadcasters with rating books, a record of who received a statement and roughly how many people believed it. AI systems don't produce that record. An output is generated once, for one person, in one session, and by default nobody -- not the claimant, not the company, often not even the AI provider -- retains a clean account of when it first appeared, how it changed across later queries, or how many other people received a version of the same false claim.

This is the gap the early cases are exposing. Showing that an AI system said something false about you is, in most instances, the easy part -- a single screenshot will usually do it. Showing that the statement reached anyone, that it was believed, and that it caused a quantifiable loss is the part that sinks the claim. And that evidence has a shelf life. AI outputs are not archived by the system for a claimant's benefit. Models get updated. Sessions expire. Answers shift from one query to the next even without a retraining event. By the time a dispute is serious enough to justify legal spend, the specific words at issue may no longer be reproducible on demand, and there may be no way to establish whether the claim appeared once or was surfacing to a meaningful share of everyone who asked.

That has a practical consequence for anyone advising a company, a public figure, or an insurer on this kind of exposure: the moment to start building a record is before there is a dispute, not after. Waiting for a complaint, a client call, or a reputational flashpoint to justify capturing what an AI system is saying means starting the evidentiary work at exactly the point when the underlying output is least likely to still exist in its original form. A false claim about embezzlement, fraud, or misconduct that appears once today and differently tomorrow needs to be caught, dated, and preserved at the moment it's first observed -- ideally across more than one AI system, since courts weighing reach and repetition will care whether a claim was isolated to one query on one system or was showing up consistently across several.

None of this requires AI companies to admit fault or change their products. It requires the people who might one day need to prove what was said, when, and how often, to stop assuming that record will simply exist when they need it. The Walters ruling wasn't a verdict on whether AI systems fabricate facts about real people -- everyone in the courtroom agreed that one did. It was a verdict on the absence of proof for everything that came after the fabrication. That is the part still fixable, and the part almost nobody is currently doing anything about.