Informational only. Not legal advice. No attorney-client relationship is created by reading this post. Consult a licensed attorney in your jurisdiction.
The doctrine of defamation law has absorbed successive disruptive communication technologies—the penny press, radio, television, the internet—without abandoning its structural logic. Generative artificial intelligence is the newest stress test, and its defining characteristic—the confident production of false statements of apparent fact—makes it a doctrinal flashpoint. The first wave of AI defamation litigation has now generated rulings, most notably the Georgia superior court's May 2025 decision in Walters v. OpenAI, and the results reveal how durable the New York Times Co. v. Sullivan framework remains—even when no human editor made the offending statement.
I. The Sullivan Constitutional Framework
New York Times Co. v. Sullivan, 376 U.S. 254 (1964), is the constitutional bedrock of American defamation law. The Court held that the First and Fourteenth Amendments prohibit a public official from recovering damages for a defamatory falsehood relating to official conduct unless the plaintiff proves the statement was made with actual malice—that is, with knowledge that the statement was false or with reckless disregard of whether it was false or not. Id. at 279–80.
Sullivan arose from a civil rights-era Alabama judgment that the Court recognized as the functional equivalent of seditious libel: a tool for suppressing criticism of government officials. The Court's answer was to constitutionalize a fault standard that privileged robust public debate, accepting that some false statements would go unremedied as the price of that privilege.
The Court subsequently extended the actual malice requirement to public figures generally—not only public officials—in Curtis Publishing Co. v. Butts, 388 U.S. 130 (1967), and further refined the public figure categories in Gertz v. Robert Welch, Inc., 418 U.S. 323 (1974). Gertz distinguished:
- All-purpose public figures: individuals of pervasive fame or notoriety who must show actual malice across all contexts;
- Limited-purpose public figures: individuals who voluntarily inject themselves into a particular public controversy, who must show actual malice with respect to statements about that controversy; and
- Private figures: individuals who retain access to the negligence standard for recovery of actual damages.
The public figure/private figure distinction is now among the most consequential doctrinal choices in any defamation case.
II. Elements of Defamation: Florida and Alabama
Florida common-law defamation requires the plaintiff to prove: (1) the defendant published a false statement of fact; (2) the statement was "of and concerning" the plaintiff; (3) the statement was communicated to a third party; (4) the statement caused injury to the plaintiff's reputation; and (5) if the plaintiff is a public figure, the defendant acted with actual malice—knowledge of falsity or reckless disregard of truth or falsity. Private figure plaintiffs in Florida need only show fault amounting to at least negligence.
Alabama common-law defamation requires the plaintiff to establish: (1) a false and defamatory statement concerning the plaintiff; (2) an unprivileged communication to a third party; (3) fault amounting to at least negligence; and (4) either actionability of the statement per se or the existence of special harm caused by the publication. See generally Restatement (Second) of Torts §§ 558–59. For public figures under Alabama law, the Sullivan actual malice standard applies. Alabama courts also recognize truth as a complete defense and absolute and qualified privileges for certain communications.
One practical note for Alabama practitioners: although Sullivan itself arose from an Alabama judgment (Sullivan was a Montgomery city commissioner suing the New York Times), the constitutional framework now preempts any state law attempt to impose liability without the appropriate fault showing for public-figure plaintiffs.
III. The AI Hallucination Problem
Large language models (LLMs) such as ChatGPT generate outputs by predicting statistically likely text sequences, not by retrieving verified facts. When a model's statistical patterns lead it to produce a confident-sounding but false assertion about a specific named individual, the resulting statement has the surface characteristics of a defamatory publication: it appears factual, it identifies a specific person, and it may be received by a human reader as a statement of fact.
This phenomenon—known as "hallucination"—is inherent to current LLM architecture. A model trained on vast text corpora may associate a person's name with concepts that appear in nearby text without having any verified factual basis for the specific assertion it produces. The result can be a fabricated legal complaint, a false allegation of financial misconduct, or an invented criminal record—all delivered in authoritative prose.
The legal question is whether traditional defamation doctrine can reach this output, and if so, who bears liability.
IV. Walters v. OpenAI: The Georgia Ruling
Walters v. OpenAI, L.L.C., No. 23-A-04860-2 (Ga. Super. Ct. Gwinnett Cnty. May 19, 2025), is currently the leading authority. Mark Walters, a nationally syndicated radio host and prominent Second Amendment commentator, sued OpenAI after ChatGPT falsely stated that he had been accused of embezzling funds from the Second Amendment Foundation (SAF) in a lawsuit with which he had no involvement. A journalist querying ChatGPT for a summary of a real SAF lawsuit received the hallucinated output; he recognized the error, verified the actual complaint, and did not publish the false information.
Judge Tracie Carson granted summary judgment for OpenAI on three independent grounds, each of which is doctrinally significant.
Ground One: No Defamatory Statement as a Matter of Law. Under Georgia law, a plaintiff must show the statement could be "reasonably understood as describing actual facts" about the plaintiff. The court found that a reasonable reader in the journalist's position—who knew ChatGPT could produce "flat-out fictional responses," had received repeated hallucination disclaimers, and had access to the actual complaint—would not have understood the output as stating actual facts without verification. Because the journalist did not in fact believe the statement, "the statement is not defamatory as a matter of law."
This holding has significant implications beyond Georgia. It establishes that AI disclaimers—if sufficiently prominent and if the recipient is an informed user—can undermine the defamatory meaning element by defeating the reasonable reader's likely interpretation.
Ground Two: No Fault, No Actual Malice. The court held Walters a limited-purpose public figure by virtue of his large radio audience (over 1.2 million listeners per segment) and his voluntary role in Second Amendment public debate. As a limited-purpose public figure, Walters was required to show actual malice—knowledge or reckless disregard of falsity—by clear and convincing evidence.
The court found no evidence of actual malice. OpenAI's unrebutted expert testimony established that the company had taken industry-leading steps to reduce hallucinations. The court explicitly rejected Walters's argument that knowingly deploying a system capable of hallucination constituted actual malice: "the mere knowledge that a mistake was possible falls far short of the requisite 'clear and convincing evidence'" of actual malice. That would amount to strict liability—a standard that has never survived Sullivan scrutiny.
The court also independently found no ordinary negligence, as Walters produced no evidence of the standard of care that a reasonable publisher in OpenAI's position would have employed.
Ground Three: No Damages. At deposition, Walters conceded he had not been damaged by the ChatGPT output and was not seeking actual damages. He was barred from punitive damages because he had not made the Georgia-required retraction request. Presumed damages were unavailable because Walters admitted he had not been harmed, rebutting any presumption.
V. Doctrinal Implications for AI Defamation Claims
Several analytical points emerge from Walters and the broader post-Sullivan framework as applied to AI.
Defamatory meaning and disclaimers. The Walters court's defamatory meaning analysis rewards developers who prominently disclaim hallucination risk. Where a reasonable user understands they are receiving potentially inaccurate output, the reasonable-reader test may fail. This creates a structural tension: the same warnings that help users should have known not to trust the output also shield developers from liability.
Public vs. private figure in AI contexts. Walters's classification as a limited-purpose public figure was central to the outcome. Had he been a private figure, the negligence standard would have applied, and the outcome might have differed. Practitioners representing private-figure plaintiffs should argue that: (a) the plaintiff never voluntarily injected himself into public controversy; and (b) even if occasionally in the public eye, the specific subject matter of the false statement falls outside any limited-purpose public figure status.
The "of and concerning" requirement. In the scenario that produced the Walters case, the hallucinated output was triggered by a query about a real lawsuit. LLMs may conflate multiple real persons or generate name-and-fact combinations that technically satisfy the "of and concerning" requirement even without a specific targeted prompt. This category presents stronger plaintiff-side arguments because the false statement's connection to the actual plaintiff is clearer.
Third-party query problem. In Walters, a third party (the journalist) queried the model. In other scenarios, the defamed individual may be the one who discovers the false output about themselves, or outputs may be incorporated into published articles. The publication-to-third-party element requires analysis of how the output reached persons other than the querying user.
Republication liability. When a news organization or individual publishes AI-generated content without verification, the republisher assumes the standard of care applicable to its own publication. A journalist who adopts AI output as their own publication—without verification—cannot hide behind the AI's disclaimers. The actual malice analysis then focuses on the journalist's knowledge, not the model's architecture.
VI. Florida and Alabama Defamation Practice Notes
Retractions. Florida's retraction statute, Fla. Stat. § 770.01 et seq., requires a plaintiff to give 5 days' written notice and an opportunity to retract before filing suit for libel against a newspaper or broadcaster. This requirement applies to "periodicals" in the traditional sense; whether AI outputs are subject to the retraction demand framework has not been litigated.
Statute of limitations. Florida: two years for libel and slander under Fla. Stat. § 95.11(4)(g). Alabama: two years under Ala. Code § 6-2-38(k). The discovery rule may apply in cases where the plaintiff did not promptly learn of the defamatory output.
Single publication rule. In internet defamation, the single publication rule typically runs the limitations period from initial publication, not from subsequent re-posting. The analysis for AI outputs that persist and are re-generated on each query is unsettled.
Identifying the publisher. Where AI output is embedded in a third-party platform (a company using an LLM API in its product), the platform may have Section 230 immunity under the Communications Decency Act, 47 U.S.C. § 230. The CDA's § 230(c)(1) immunizes "interactive computer service providers" for third-party content. Whether AI-generated output is "third-party content" for § 230 purposes is a live question; the argument that the developer is itself the content creator (and thus not immune) has considerable traction.
VII. Open Questions
The Walters decision leaves several critical issues unresolved. First, the opinion rested in part on the fact that the journalist did not actually publish the false statement and admitted disbelief—facts that will not recur in every case. Second, the decision provides limited guidance on private-figure plaintiffs, who need only show negligence. Third, the analysis of AI disclaimers as negating defamatory meaning may not hold as AI output becomes more authoritative and less hedged.
The deepest doctrinal question is whether Sullivan's actual malice standard—designed to protect deliberate editorial choices—should apply when a model produces output without any human scienter about the specific statement. Several scholars argue that the standard must be reconsidered for AI; courts have so far applied it as written.
Conclusion
Sullivan and its progeny remain the governing framework for AI defamation claims. The Walters decision demonstrates that defendants who invest in hallucination mitigation and disclosure can mount powerful defenses at the summary judgment stage—even when the challenged statement is objectively false. For plaintiffs, the path runs through private-figure status where available, documented actual harm, republication by a human publisher, and novel arguments about the limits of AI disclaimers as a liability shield.
Talk to Yates Anderson
If you are litigating a matter in this area — or weighing whether to — the working analysis above only goes so far. Request a case evaluation and a Yates Anderson attorney will respond within one business day.
Informational only. Not legal advice. No attorney-client relationship is created by reading this post. Consult a licensed attorney in your jurisdiction.