AI defamation: who is liable when AI output lies about you

The short version

• Defamation is a false statement of fact, published to someone else, that identifies and harms a person. The same elements decide the claim whether a person or an AI tool produced the statement.
• You sue a person or a company, not the AI. A model has no state of mind, and fault is measured against the people who built or used it. You cannot sue ChatGPT itself.
• The only US court to decide one so far, in Walters v. OpenAI, granted OpenAI summary judgment on three independent grounds: no defamatory meaning as a matter of law, no fault, and no damages.
• Section 230 of the Communications Decency Act stops a service being treated as the publisher of information provided by “another information content provider”. Whether that covers a model’s own output is unsettled, and Walters did not decide it.
• Little is settled: one trial-level state ruling, one case sent to arbitration, and no appellate decision, as of 11 November 2026.
• The general question of who is liable for what AI says sits in AI in law; lawyers sanctioned for filing citations an AI invented sit in AI hallucinations and court sanctions.^1,2,4

What are the elements of defamation, and do they apply to AI output?

Yes, the ordinary elements apply. A defamation claim needs a false statement of fact, published to a third party, that is “of and concerning” an identifiable person, made with the required fault, and causing damage. A generative AI tool, software that writes new text, changes who is in the frame, not the test.²

Defamation is a false statement of fact that harms someone’s reputation. Written down it is libel; spoken, it is slander. Because AI output is written, an AI defamation claim is almost always one for libel. Publication means the statement reached at least one person other than the subject. A hallucination is a confident AI output that is not true, and a hallucination about a named person is where an AI defamation claim starts.²

The first element did the most work in Walters. Georgia law asks whether a statement can be reasonably understood as describing actual facts about the plaintiff, judged not by one misled reader but by an objective, hypothetical reasonable reader, with any warning on the tool counted as one factor. That is the standard the order applied to a ChatGPT output.²

Identification, often called “of and concerning”, asks whether the output points to a real, identifiable person. Battle v. Microsoft turns on it. The memorandum opinion, describing what the plaintiff alleges, records that results from Microsoft’s Bing and its Copilot chatbot “conflat[e] Mr. Battle with a person of a similar name” who the complaint says is a convicted terrorist. That is an allegation, not a finding; the court did not rule on it.^3,4

What did the only US court to decide an AI-defamation case hold?

In Walters v. OpenAI (Superior Court of Gwinnett County, Georgia, 19 May 2025), the court granted OpenAI summary judgment, a ruling that ends a claim before trial because the key facts are not in genuine dispute, on three independent grounds, so OpenAI won without any new AI rule. The answer to whether OpenAI won the ChatGPT defamation case is yes, on ordinary defamation law.²

The order records the facts. A journalist asked ChatGPT to summarise a real complaint in another lawsuit. The output falsely said the complaint accused Mark Walters, a radio host, of embezzling from a foundation. The journalist did not believe it, confirmed within about ninety minutes that it was untrue, and did not pass it on.²

First, meaning. The court held that a reasonable reader in the journalist’s position could not have concluded that the output communicated actual facts, so it was not defamatory as a matter of law. The court did not create a disclaimer safe harbour. The order reproduces the on-screen notice the journalist saw, that “ChatGPT may produce inaccurate information about people, places, or facts”, and weighed it as one factor, alongside the reader’s own doubt, his experience of past fabricated outputs, and his access to the real complaint.²

Second, fault. A defamation plaintiff must show at least ordinary negligence. The court rejected the argument that a provider is negligent merely because it knows the model can make a mistake; that would impose strict liability, liability without fault, which Georgia and constitutional law forbid, the rule drawn from Gertz v. Robert Welch, Inc. A private figure needs only negligence. As a public figure, someone in the public eye for the matter in dispute, Walters had to show actual malice, which the court defined as a subjective awareness of probable falsity, the standard from New York Times Co. v. Sullivan. That had to be shown by clear and convincing evidence, a high burden, and Walters had no evidence of it.²

Third, damages. Walters conceded he had no actual damages. Punitive damages were barred because Georgia law requires a libel plaintiff to ask for a correction or retraction first, which he did not do. Any presumed damage was rebutted on his own admissions.²

The reach is narrow. This is a trial-level state order, signed on a proposed order the court edited, and tied closely to its facts. It is not binding precedent. The motion to dismiss had been denied earlier; this was the later summary-judgment stage. As of 11 November 2026 we found no appeal.²

Who can you sue: the AI, the developer, or whoever repeats the output?

Not the AI itself. A claim runs to a person or a company: the developer that operates the model, or a user who publishes the output onward. Fault and state of mind attach to people, not to software, so the question of whether the AI developer or the user is liable depends on who did what.²

The developer can be sued directly. In Walters the developer was sued directly and the claim was resolved on the ordinary defamation elements, not on any immunity, and the developer prevailed. The court measured fault by what a reasonable publisher in the provider’s position would have done, not by whether the output came from a large language model.²

A republisher is the clearer target. Under ordinary defamation law, anyone who repeats an AI output as fact can face republication liability, the rule that each fresh publication of a defamatory statement is its own wrong. That rule is general defamation law, not part of the Walters order; in Walters the journalist avoided it by not passing the output on. An AI chatbot can produce a defamatory statement, but it is the person who passes it on as true who is easiest to hold answerable.²

The route to a decision matters. In Battle v. Microsoft the court never reached the merits. On 23 October 2024 it granted Microsoft’s motion to compel arbitration and stayed the case, so the AI-liability questions stayed unanswered. The matter is undecided as of 11 November 2026.⁴

Does Section 230 shield AI companies for what their chatbot says?

Unsettled. Section 230(c)(1) says “No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information” that is “provided by another information content provider”. The shield turns on that word “another”.¹

Section 230, part of the Communications Decency Act, was enacted in 1996 and last amended in 2018. It was written for third-party content that a service hosts, such as a user’s post, not for text a model generates itself. The statute says nothing about generative AI.¹

The hinge is a definition. Section 230(f)(3) defines an “information content provider”, the party whose content is not shielded, as anyone “responsible, in whole or in part, for the creation or development of information”. If a provider itself created or developed the output, (c)(1) may not cover it. The text does not resolve this, which is why whether Section 230 protects generative AI output is a live debate.¹

Walters did not settle it. The court ruled on meaning, fault and damages and did not reach Section 230. No decision we found resolves whether Section 230(c)(1) reaches a model’s own generated output.^1,2

What can you do if an AI invents something false about you?

This is practical, not legal advice. If an AI output names you falsely, keep a copy of the exact output with the date and the prompt, and send the provider a notice asking it to correct or retract the statement, which is how you ask an AI company to remove a false statement about you. In Walters, Georgia’s retraction requirement decided the punitive-damages question, and the plaintiff in Battle alleges he told the provider. A dated correction request is on the record and costs little.^2,3

What is still undecided in AI-defamation law?

Almost everything of substance. The record so far is a single trial-level order for the defendant and one claim pushed into private arbitration, with no appeal decided.^2,4

• Whether an AI output is a “statement of fact” at all, or non-actionable where the reader is warned it can be wrong.
• Which fault standard fits an AI developer, and what the ordinary-negligence test requires of one.
• How a private-figure plaintiff, who needs only negligence, or an output seen by many rather than by a single reader, would fare.
• Whether Section 230(c)(1) reaches a model’s own generated output.
• What any arbitration in Battle decides, which would be private and would set no precedent.
• No binding appellate precedent on AI defamation yet exists.^1,2,4

What does this mean for anyone building or relying on AI?

In our view, existing defamation law already reaches what a generative AI system produces. No special AI rule was needed for OpenAI to win in Walters, and none has been written. A person or a company stays accountable for what its AI puts out. The model is not a defendant, and it is not an excuse. Because the ordinary elements already decide the claim, check an AI output before relying on it or repeating it as fact, and treat a provider’s warning as a factor a court weighs, not a shield. The pace of AI is ahead of the settled case law. One trial-level order and one case sent to arbitration do not resolve who answers when a model defames a widely known person, or whether Section 230 covers a model’s own words. The safest reading is that the ordinary duties govern now, and the cases and the Section 230 question are filling in the detail. For related ground, see AI governance standards and AI washing and false AI claims.^1,2