An AI said it told 2,843,917 people.
Two people asked him if it was true.
On Friday a Delaware judge refused to throw out Robby Starbuck’s defamation suit against Google, finding it sufficiently pleaded that the challenged outputs reached third parties at all. The allegation the court declined to treat as impossible: that Google’s own AI, asked directly, disclosed it had produced false statements about him to 2,843,917 unique users.
Two people came to Robby Starbuck and asked him whether something was true. A Google AI product had described him, and neither, the court noted, discounted what it said out of hand — both wanted to hear from him first. Each approach came in 2025, on the court’s reading of the complaint now pending in Delaware Superior Court — well over a year after he had found it himself, in December 2023. When he put the question to the system directly, it answered that it had been conveying falsehoods about him since October of that year: roughly two months of description he had no way to see while it was happening.
Then came a number. Google’s AI, the complaint alleges, revealed that it had produced false statements about Starbuck to 2,843,917 unique users. Google’s answer was that this admission cannot be trusted: Starbuck’s own case rests on the claim that the AI routinely misrepresents him, so he cannot turn around and ask a court to believe it. Judge Adams declined to draw that conclusion. Doing so, she wrote, would require the court to infer that the AI never told the truth — an inference that runs toward the moving party, which is not what a motion to dismiss permits. On this record the court declined to find the allegation, in its words, “so totally incredible as to be disregarded.”
On July 24, the court denied Google’s motion to dismiss. The court found the publication element sufficiently pleaded, rejecting the argument that Starbuck had to name specific recipients — a requirement that would effectively close the courthouse to plaintiffs in mass-publication cases. It set Google’s disclaimer defense aside for a fuller record rather than resolving it. And it let an actual-malice theory proceed on the possibility that people responsible for policing the outputs knew and took no corrective action while publication continued. The court was explicit that the motion did not force it into the novel questions, and resolved it on established defamation caselaw and Delaware’s plaintiff-friendly pleading standard. No finding of falsity. No finding of liability.
Edition No. 020 filed a Munich court holding that an AI’s answer is the platform’s own speech. Delaware got to the same place by a different route — Munich asked whose words those are, Delaware asked whether anyone received them. Asking who received it forces a distribution record into view. That record is worth more than the liability question that summoned it.
What the filing actually exposes
Notification is not a feature anyone built. Word of what a system says about you reaches you socially or not at all — a colleague, a client, a stranger who was unsure enough to check. Everyone else who was told simply believes it and moves on.
The count is claimed; the report does not exist. A seven-digit figure implies the distribution is tracked somewhere. The court held only that the claim survives a motion to dismiss — not that anything inside the system is counting. Whatever is kept faces inward. In this case the number surfaced two ways only: the complaint says the machine produced it, and a lawsuit put discovery within reach. Neither is an audit. Each one required suspecting first.
Correction has no address. To reach anyone about it, Starbuck posted publicly and tagged the chief executive and a chief scientist by name. An employee replied the same day, and the correspondence continued. That is the state of the art in getting a machine to stop: a public post, aimed at executives, and the goodwill of one employee who, by her own account, tried and did not succeed.
What a courtroom can’t reach
The door Friday opened is smaller than it looks, and the reason sits in the elements. Defamation requires a false statement of fact about an identifiable person that reached somebody else; because the parties agreed for this motion that Starbuck is a public figure, it also requires proving actual malice. Every one of those elements presumes the machine had something specific to say. Starbuck could sue because the system held a definite, detailed classification of him — one he alleges was fabricated outright, and one Google did not argue was true. That is authority misclassification in its most extreme and most legible form, and legibility is precisely what made it actionable.
The ordinary version runs the other way. An answer engine can leave a firm out of the answer, describe it in terms too vague to act on, or file it under a category that costs it every comparison it enters. None of that is a false statement of fact about anyone. There is no filing to make. An omission has no plaintiff, and nobody comes to ask whether the answer that never mentioned you was true. Digital Derangement Syndrome™ is that condition — the recognition failure that leaves no evidence behind — and the correction for it has never lived in a courtroom. Answer Engine Authority™ corrects it upstream of distribution, because contesting a classification means contesting a copy of it that several million people already hold.
All procedural facts, allegations and quotations are drawn from the court’s own opinion: Starbuck v. Google LLC, C.A. No. N25C-10-211 MAA, Superior Court of the State of Delaware, Adams, J., submitted April 6, 2026, decided July 24, 2026 — slip opinion (PDF, as posted by Reason). The court denied the motion to dismiss under Delaware’s reasonable-conceivability standard; it made no finding that any statement is false and no finding of liability, and expressly deferred the disclaimer questions to a later stage. In the court’s own words, it “convey[s] no agreement with the truth of the matters asserted in the Complaint,” and notes it will not necessarily repeat the word “alleged” throughout — a pleading-stage convention, not a finding.
The 2,843,917 figure originates in the complaint, not in any court finding — Starbuck v. Google LLC, complaint filed October 22, 2025 (PDF). The complaint was read in full for this edition; characterizations of what the two individuals said are the court’s, drawn from its summary of paragraphs 92 and 93.
Coverage of the ruling — The Volokh Conspiracy (Reason), July 24, 2026, and Bloomberg Law, July 24, 2026.
The Munich ruling referenced above was filed in full, with its primary sourcing, in Edition No. 020.
This is analysis of a public court record, not legal advice. The case is at the pleading stage. Nothing in it has been decided.