ChatGPT accused Mark Walters of embezzling millions from a group he never worked for

A reporter asked for a case summary and got an invented crime

In early May 2023, Fred Riehl, a journalist who edits the firearms site AmmoLand and then sat on the board of the Second Amendment Foundation, was looking into a real lawsuit his group had just filed against Washington State's attorney general. He asked ChatGPT to summarize the complaint and gave it the link.¹

The summary ChatGPT produced described a different lawsuit altogether. It said the complaint had been filed against Mark Walters, a Georgia radio host, whom it called the foundation's "treasurer and chief financial officer" and accused of misappropriating its funds. Walters held neither job and had no relationship with the foundation.¹

When Riehl asked to see the full text, ChatGPT produced one. It was a complete, correctly formatted complaint, captioned as a real federal filing with a court and a case number. The document alleged that Walters had embezzled foundation funds "in excess of $5,000,000," and it did not exist. Riehl checked with the foundation's head and confirmed within about an hour and a half that the output was false. He never published it.¹²

Walters was not the first person ChatGPT accused

Weeks earlier, the same system had done this to other people. It named Jonathan Turley, a law professor at George Washington University, as an example of a professor accused of sexual harassment, saying he had made suggestive comments to a student on a class trip to Alaska and citing a Washington Post article dated March 21, 2018. There had been no trip, no such article, and no such accusation, and the chatbot had placed the episode at Georgetown, where Turley has never taught. Another law professor, Eugene Volokh, surfaced the output by asking ChatGPT for examples of harassment by law professors.³⁴ Turley did not sue.⁴

Brian Hood, the mayor of Hepburn Shire in Victoria, Australia, was described by ChatGPT as a man convicted and jailed in a foreign-bribery scandal. In fact Hood was the whistleblower who had reported that bribery to the authorities, and he was never charged. His lawyers sent OpenAI a formal concerns notice in 2023, but he never filed suit and dropped the matter by February 2024, saying the cost of a Supreme Court defamation case was prohibitive for an ordinary person. By then OpenAI had changed ChatGPT so that it no longer repeated the claim.⁵

In March 2025, a private Norwegian citizen, Arve Hjalmar Holmen, asked ChatGPT who he was. It answered that he had been convicted of murdering two of his sons, had attempted to murder the third, and was sentenced to 21 years in prison. Holmen has three sons and no criminal record. A European privacy group, noyb, complained to Norway's data-protection authority, not for defamation but under the rule that personal data must be accurate.⁶

These four cases look alike, but only Walters filed a lawsuit and got a decision on the merits. Turley chose not to sue. Hood withdrew before he ever filed. Holmen's complaint is a data-protection matter, not a libel claim at all. So the question of civil liability is not only unsettled. It is largely untested, because most of the people a chatbot names falsely do not, or cannot, take it to court.

A chatbot libels a real person as easily as it praises one

Each of these outputs was a hallucination: text generated to read plausibly, with nothing in the system checking it against the facts it asserts. A separate lesson explains why a model writes a convincing false reference the same way it writes a real one. A lawyer who filed six court cases that did not exist, or a consultancy whose paid report cited books no one had written, met the same failure aimed at documents. Here it is aimed at a person.

A model builds each sentence by predicting a likely next word, and nothing in it registers that "the treasurer embezzled the foundation's funds" pins a crime on a living man while "the treasurer kept the foundation's books" does not. It wrote that Walters embezzled millions¹ and that Holmen killed his children⁶ with the same even confidence it would use to get a birthday right.

That is why the harm is a different kind. A false legal citation wastes a court's time and embarrasses the lawyer who filed it. A false statement of fact about a named person, delivered to someone else, is the raw material of defamation, and it does the damage without anyone having to act on it. Being described to even one reader as an embezzler or a child-killer is itself the injury, and a chatbot can produce that injury in the tone of a reference work millions of people trust.

The ruling that cleared OpenAI decided less than it seems

Walters sued OpenAI in Georgia in 2023. A judge refused OpenAI's early request to throw the case out, so the claim was not frivolous on its face.⁷ By OpenAI's own lawyers' account, it is the first defamation case in the United States against an AI company over something its chatbot made up.⁸

In May 2025, Judge Tracie Cason of the Superior Court of Gwinnett County granted OpenAI summary judgment. She rested the decision on three separate grounds, each by itself enough to defeat the claim.²

The first was meaning. A statement is defamatory under Georgia law only if a reasonable reader would take it as stating actual facts about the plaintiff. The court held that no reasonable reader in Riehl's position could have. OpenAI's warnings told users the model can produce inaccurate output, and Riehl had not believed it in any case.

The second was fault. A plaintiff must show the publisher was at least careless. Walters's lawyers argued OpenAI was negligent simply for releasing a system that "makes up random false statements," but the court held that this would amount to strict liability, which defamation law does not permit. It also found Walters a public figure, a prominent voice in the gun-rights debate who says he reaches more than a million listeners per segment, so he had to prove actual malice, and could not.

The third was damages. Walters conceded in his deposition that he was claiming no harm and sought none. He had also never asked OpenAI to correct or retract the output, which Georgia law requires before a plaintiff can seek punitive damages. The court found he could recover nothing.²

OpenAI's own lawyers drafted the order, and the judge signed it "as edited by the court." That is routine in civil practice, where the winning side often writes the proposed order, but it means the reasoning is the prevailing party's, adopted rather than composed by the court.² OpenAI's counsel said the decision "will help guide future litigation."⁸

What the ruling did not do is decide whether an AI company can be liable for defamation. Each of its three grounds is tied to the facts of Walters's case: a sophisticated reader who never believed the output, a plaintiff who claimed no damages, a plaintiff who never asked for a retraction. Change those facts and the analysis changes. A brief filed in the case pressed exactly that point.

Volokh argues that because OpenAI presents ChatGPT as a reliable source, its statements should count as assertions of fact. RonNell Andersen Jones, a University of Utah law professor, suggests the better legal fit for this harm may be product-liability law rather than defamation at all.⁴

The assistants doing this now arrive with citations that do not exist

The setup that produced the Walters output is standard and still running. A general-purpose assistant, asked about a person, generates a fluent biography whether or not it has real information, and nothing in it checks the result against the record. Systems that search the web first reduce this but do not remove it. The clinics' brief noted that ChatGPT had passed 200 million weekly users and that OpenAI's tools held roughly three-quarters of the market, figures it drew from outside reports.⁹ The false claims now often arrive dressed in invented citations: article titles, dates, and case numbers. That is what made ChatGPT's nonexistent Washington Post article and its fabricated federal complaint read as real.³¹

Most people a chatbot names falsely never go to court, which is why the question stays untested. Walters reached a judgment, and it answered only the narrow questions his own case raised. For the next person a chatbot accuses of a crime, the law's answer to who is responsible is mostly unwritten.²