The Ruling That Turned Correction Emails Into Evidence

2026-09-25

On September 14, a federal judge in Chicago refused to throw out a defamation claim against Google over its AI Overviews. The headline write-ups focused on what the AI system got wrong: it told searchers that James Keene, the author whose story became the Apple TV series Black Bird, was serving a life sentence for murdering three women. He was not. Keene was convicted of a drug conspiracy in 1997, later worked as an FBI operative, and had that conviction expunged.

The error is the least interesting part of the ruling. AI systems generate false statements about real people constantly, and courts have mostly treated that as an unfortunate but unremarkable fact of how these tools work. What made Judge Thomas Durkin's opinion different is that he did not rule on whether the statement was false. He ruled on what Google knew, and when.

Keene had notified Google three times -- on May 27, 2025, again on May 31, and once more after the same false claim resurfaced. To clear the bar for defamation as a public figure, Keene had to show Google acted with actual malice: publishing a false statement knowing it was false, or with reckless disregard for whether it was true. Durkin found the repeated notices did the work. Either the reports reached someone at Google who knew the statement was false and let it stand, or nobody responsible reviewed a legal removal request at all. Both possibilities, at this stage, were enough to let the claim proceed.

Two of Keene's four Overview claims survived the motion to dismiss. Two were dismissed without prejudice. Google has not been found liable for anything -- this is a pleading-stage ruling, not a verdict. But the reasoning is the part worth sitting with, because it reframes what a correction request is for.

Before this ruling, a notice to a company that its AI system was publishing something false functioned mostly as a request: please fix this. After it, that notice starts to function as a timestamp: on this date, this entity was told this exact thing was false. What happens between that date and the next time someone checks becomes the substance of the legal claim. Not the falsehood itself -- the interval during which it was known to be false and kept circulating anyway.

That shifts the practical question for anyone managing a company's, a client's, or a country's exposure to AI-generated claims. It is no longer just "what is AI saying about us." It is "can we show, with a date attached, exactly what was said, that we flagged it, and what was still being said afterward." A screenshot taken once, sitting in someone's downloads folder, does not answer that question well. Neither does a vague recollection that "the AI said something wrong a few months ago." The Keene opinion turned on specifics -- three dated notices, a description of what changed and what didn't -- because specifics are what a court needs to distinguish an honest mistake from a pattern a company chose to ignore.

There's a structural reason this keeps happening: AI outputs are not static pages. What a system says about a person or a company today can be different tomorrow, different again after a model update, and different depending on which system answers. A single screenshot captures a moment, not a pattern, and a pattern is what actual-malice claims are built on. Reconstructing that pattern after the fact -- pulling old queries, guessing at dates, hoping a cached version still exists somewhere -- is exactly the position Keene's lawyers were not in, because he had emailed Google as it was happening.

The Munich case earlier this year pointed the same direction from a different legal tradition: once a company is on notice that an AI output about someone is false, continuing to serve it becomes the company's own statement to answer for, not an accident to shrug off. Regulators and courts on both sides of the Atlantic are converging on the same instinct -- knowledge plus inaction is where liability lives.

None of this requires AI systems to become accurate, which none of them will be anytime soon. It requires being able to prove, later, what a system said, when it said it, and who was told. That's a different kind of readiness than most reputation and legal teams have built for. The AI Overview that misstates a fact today is a nuisance. The same output, undocumented, unflagged, and unaddressed for two months, is a fact pattern.