The Number That Got Google's AI Into Court Came From the AI Itself

2026-07-30

On July 24, a Delaware judge did something no US court had done before: she let a defamation claim over a chatbot's own output move into discovery.

The case is Robby Starbuck's lawsuit against Google, filed last October over claims that Bard, Gemini and Gemma told users he was a rapist, a child abuser, and an associate of Jeffrey Epstein, on top of inventing a criminal record involving stalking and drug charges. Google asked the court to throw it out, leaning on a defense that has worked before: these were hallucinations, not statements, and a system that doesn't know what's true can't defame anyone. Judge Meghan Adams disagreed. Starbuck, she wrote, "pleads sufficient facts indicating the Outputs were published to third parties" — enough to proceed toward trial.

That single finding is the whole story, because it's exactly where the last comparable case died. In May 2025, a Georgia court threw out radio host Mark Walters' defamation claim against OpenAI for almost the identical reason, in reverse: the only person who ever read the false statement about him was a journalist testing the system, who recognized it as untrue and never repeated it. No second reader, no evidence anyone's opinion of Walters changed, nothing to compensate. Same underlying problem — a chatbot stating a specific, damaging falsehood as fact — and two opposite outcomes, turning entirely on whether anyone could show the statement actually reached other people.

So what did Starbuck have that Walters didn't? A number. His complaint alleges that when asked, Gemini itself reported that the fabricated claims about him had been shown to 2,843,917 unique users.

Sit with where that figure comes from. It isn't a server log Google produced. It isn't an independent count from a monitoring service. It's a number the same system generated when someone asked it — the same system whose entire legal defense in this case is that it invents specific, confident-sounding facts it has no way of actually knowing. A system capable of fabricating a criminal record and an Epstein connection is, on the current record, also the source for the one figure now doing the legal work of turning "an AI said something false" into "a statement was published to the public." Nobody has yet had to test whether a chatbot's self-reported usage count holds up, because until three weeks ago no case had gotten far enough to need one. Discovery is where that test happens. If the reach figure was generated the same way the underlying allegations were, the case that just cleared a motion to dismiss could run into a different problem entirely.

That irony is worth sitting with, but it's not really the point. The point is what the ruling establishes going forward, regardless of how the reach number holds up: publication and repetition are becoming the pivot on which these cases turn. Not whether an AI system said something false about someone — courts on both sides of the Walters and Starbuck rulings agree that part happens routinely. What decides the case is whether anyone can show the statement reached people, more than once, in a form a court will accept as evidence rather than argument.

And almost nobody outside an active lawsuit has a way to establish that independently. Most organizations and individuals learn what an AI system has said about them the way Starbuck did originally — because someone happened to ask, took a screenshot, and sent it along. That tells you the claim existed once, for one person, on one day. It tells you nothing about whether it was an isolated fluke or a pattern the system was repeating to everyone who asked a similar question that week, across however many systems people were actually using. The only account of scale available to Starbuck's legal team was the defendant's own product, asked to grade its own homework.

That's the gap this ruling puts a spotlight on without resolving it. A court has now said that reach and repetition matter enough to get a case to discovery. It hasn't said what counts as adequate proof of reach when the only available account is generated by the same system accused of fabricating the underlying claim in the first place. Until that gets settled, the safest position for anyone who might one day need to show what an AI system said about them — and how often, and to how many people, and starting when — is not to wait for the system to volunteer that account after a dispute has already started. By then, as the Delaware and Georgia rulings both show in their own way, whoever built the record first decides what the case actually gets to argue about.