Walters v. OpenAI: Right of Publicity

A Georgia court granted OpenAI summary judgment over a ChatGPT hallucination, holding no reasonable reader would take the fabricated output as fact.

In Walters v. OpenAI, L.L.C., No. 23-A-04860-2 (Ga. Super. Ct., Gwinnett County, May 19, 2025), a Georgia trial court issued what legal commentators describe as the first-of-its-kind merits ruling on whether a generative-AI system can be liable for defaming a real, identifiable person through fabricated output. Mark Walters, a nationally syndicated radio host, sued OpenAI after a journalist prompted ChatGPT to summarize a federal lawsuit and the model produced a fabricated account stating that Walters had embezzled funds from the Second Amendment Foundation. Walters was not a party to that lawsuit, and the embezzlement claim was entirely false, a so-called hallucination. Judge Tracie Cason granted OpenAI summary judgment, holding on three independent grounds that Walters could not make out a defamation claim.

The decision sits at the doctrinal frontier where defamation law meets generative AI, adjoining the digital-replica and right-of-publicity questions that arise when machines generate statements and images about real people. Because it is among the first substantive rulings on AI-generated falsehoods about identifiable individuals, it offers an early map of how courts may analyze such claims.

At a glance

Case: Walters v. OpenAI, L.L.C., No. 23-A-04860-2 (Ga. Super. Ct., Gwinnett County).
Decided: May 19, 2025 by Judge Tracie Cason; summary judgment granted for OpenAI on the defamation claim.
Holding: ChatGPT’s fabricated statement was not actionable because no reasonable reader would have taken it as fact, the plaintiff (a public figure) could show neither actual malice nor negligence, and he proved no damages.
Status: Trial-court decision as of July 2026; no appeal to the Georgia Court of Appeals appears on the public record.

The doctrinal frame: defamation and the reasonable reader

Defamation requires a false statement of fact about the plaintiff, published to a third party, made with the requisite degree of fault, and causing damages. A threshold question is whether the challenged statement conveys a provably false assertion of fact rather than opinion, rhetoric, or something a reasonable audience would not take literally. Statements that no reasonable reader would understand as stating actual facts are not actionable, a principle that does substantial work in the AI context because model outputs come wrapped in disclaimers and are widely known to be unreliable.

Fault is the second axis. Under New York Times Co. v. Sullivan and its progeny, a public figure must prove actual malice, meaning knowledge of falsity or reckless disregard for the truth, by clear and convincing evidence. A private figure generally must show at least negligence. Damages are the third axis. Some categories of defamation historically permitted presumed damages, but constitutional doctrine constrains when presumed and punitive damages are available, and Georgia’s retraction statute separately bars punitive damages where the plaintiff never requested a correction.

The facts and the hallucinated summary

A journalist, Frederick Riehl, was researching a real federal lawsuit and asked ChatGPT to summarize it. The model produced a response asserting that the suit involved Mark Walters and accused him of embezzling money from the Second Amendment Foundation, where it described him as chief financial officer. None of that was true. Walters was not named in the underlying case, held no such position, and had not been accused of embezzlement. The output existed only in the exchange with Riehl.

Critically for the analysis, Riehl did not believe the output. He was aware that ChatGPT can generate inaccurate information, the model had indicated it could not access the internet or the linked complaint, and he confirmed the falsity within roughly ninety minutes. He never published or repeated the statement. Walters nonetheless sued OpenAI for defamation.

The holding: three independent grounds

Judge Cason granted OpenAI summary judgment on three separate and independently sufficient bases.

First, the output was not defamatory because no reasonable reader in Riehl’s position would have understood it as stating actual facts about Walters. The court pointed to the surrounding context: OpenAI’s terms warned that outputs may be inaccurate, ChatGPT itself signaled its inability to access the source document, the claim was implausible on its face given the recipient’s knowledge, and Riehl in fact recognized and confirmed the error quickly without relying on it. In that setting, the statement functioned not as an assertion of fact but as an evidently unreliable machine output.

Second, Walters could not establish the required fault. The court treated him as a public figure, subject to the actual-malice standard, and found no clear and convincing evidence that OpenAI acted with knowledge of falsity or reckless disregard for the truth. To the contrary, the record showed OpenAI had taken extensive measures to reduce hallucinations. The court went further, concluding that Walters could not show even ordinary negligence, because he offered no evidence of the standard of care a reasonable publisher in OpenAI’s position would owe or how OpenAI breached it.

Third, Walters proved no damages. He conceded he had suffered no actual harm and sought no actual damages, and the output had reached only Riehl, who never believed or republished it. The court rejected presumed damages on two grounds: any presumption of reputational harm was rebutted by Walters’ own admissions, and the First Amendment bars presumed damages absent actual malice where the statement concerns a matter of public concern. It separately barred punitive damages under O.C.G.A. § 51-5-2, because Walters had not requested a correction or retraction as Georgia law requires before punitive damages may be sought.

What the decision signals and what it leaves open

The ruling is significant less for any sweeping rule than for how it applies settled defamation doctrine to a new technology. Each of the three grounds is a familiar defamation hurdle, and the opinion shows that those hurdles can be formidable for plaintiffs suing over AI hallucinations, especially where the output is not believed, not republished, and causes no harm. It does not hold that AI output can never be defamatory. It holds that on this record, against a public-figure plaintiff who suffered no damages, the claim failed.

The decision also illuminates the adjacent terrain of identity and publicity. Generative systems increasingly produce statements and images about real people, raising overlapping questions of defamation, false light, and the right of publicity. Walters suggests that context and audience understanding will be central to all of them, and that plaintiffs will need to show both that a reasonable audience took the output as fact and that concrete harm followed.

Open questions

What about a believed and republished output? The result rested heavily on Riehl’s disbelief and non-republication. A recipient who credited and spread a hallucinated accusation could present a materially different case.
How are private-figure plaintiffs treated? The court applied the public-figure standard to Walters. The negligence analysis for a private individual defamed by AI output remains largely untested.
What is the standard of care for an AI developer? The court found no evidence of a reasonable-publisher standard OpenAI breached, but did not define what such a standard requires, leaving the negligence framework for AI outputs unsettled.
How will damages be shown? Walters sought no actual damages. Future plaintiffs will have to establish measurable reputational or economic harm from ephemeral, single-recipient outputs, a difficult evidentiary task.

Implications for creators and businesses

Context and disclaimers matter. The presence of accuracy warnings and the audience’s awareness of AI unreliability weighed heavily against liability. Clear disclosure of a system’s limitations is both a design choice and a legal one.
Republication is where risk concentrates. The output here harmed no one because it was not believed or spread. Businesses that take AI output and publish it as fact assume the defamation risk the developer avoided.
Public-figure status raises the bar. Actual malice is hard to prove against a developer that documents its efforts to reduce hallucinations, which favors AI companies in suits by public figures.
Damages must be real. Proof of concrete harm is essential. A presumption of reputational harm can be rebutted by the plaintiff’s own admissions, and in Georgia a libel plaintiff who never requests a correction or retraction is statutorily barred from recovering punitive damages.

Frequently asked questions

What was Walters v. OpenAI about? Radio host Mark Walters sued OpenAI after ChatGPT generated a false summary of a real lawsuit, stating that Walters had embezzled funds from the Second Amendment Foundation. Walters was not a party to that suit and the claim was fabricated. He sued for defamation. The Georgia Superior Court in Gwinnett County granted OpenAI summary judgment on May 19, 2025, in what firms describe as the first-of-its-kind merits ruling on defamation by generative-AI output.

Why did OpenAI win? The court gave three independent reasons. First, no reasonable reader in the recipient’s position would have understood the output as stating actual facts, given disclaimers, the recipient’s knowledge that ChatGPT can be inaccurate, and his quick confirmation of the falsity. Second, Walters, treated as a public figure, could not show actual malice, and could not even show ordinary negligence. Third, he sought and proved no actual damages, and could not recover presumed or punitive damages.

Does Walters mean AI companies can never be liable for false outputs? No. The ruling is a trial-court decision on a specific record, and it turned heavily on context: the recipient did not believe the output, did not republish it, and suffered no harm, and the plaintiff was a public figure who requested no correction. A different plaintiff, a believed and republished falsehood, or proof of actual damages could produce a different result. The decision maps the hurdles rather than granting blanket immunity.