Large language models fabricate. They have done so since deployment, and they will continue to do so. When the fabrication identifies a real person and attributes false criminal conduct, financial fraud, or other character-destroying accusations to that person, we have a classic defamatory statement—embedded in a technological delivery mechanism that has no historical parallel. The first wave of AI defamation litigation has produced initial rulings, some favorable to defendants, that frame the doctrinal questions clearly. The practitioner evaluating an AI defamation claim today must navigate four interlocking issues: who published the statement, what standard of fault applies, whether the hallucination constitutes a statement of verifiable fact, and how the single-publication rule interacts with the architecture of LLMs.
The Cases So Far
Walters v. OpenAI, No. 23-A-04860-2 (Ga. Super. Ct., Gwinnett County)
In June 2023, radio host Mark Walters filed what is recognized as the first major AI defamation lawsuit against an American AI company. A journalist had prompted ChatGPT to summarize a lawsuit filed by the Second Amendment Foundation. ChatGPT responded with a detailed, entirely fabricated account in which Walters—who had no connection to the case whatsoever—was accused of embezzling funds, manipulating financial records, and failing to provide accurate reporting. None of this was true; Walters was never a party to or named in any such lawsuit.
The case proceeded through discovery. OpenAI moved for summary judgment in early 2025, and on May 19, 2025, the Gwinnett County Superior Court granted summary judgment in OpenAI's favor on three independent grounds: (1) the ChatGPT output could not be "reasonably understood as describing actual facts" given that the chatbot had itself warned the journalist it lacked access to the requested document, OpenAI's terms of service warned of potential inaccuracies, and the journalist quickly recognized the output as fabricated; (2) Walters failed to demonstrate the requisite fault—OpenAI neither knew the output was false nor acted with reckless disregard for its truth or falsity; and (3) Walters had not suffered compensable damages on the record presented.
Walters is a significant early data point but not a categorical holding. The court's reasoning was fact-specific: the disclosure of uncertainty by ChatGPT itself, the journalist's own rapid recognition of the fabrication, and the absence of third-party transmission of the false content all worked against the plaintiff. Cases involving cleaner fact patterns—where the hallucination was transmitted to third parties who relied on it, where the AI did not disclaim uncertainty, or where the plaintiff can demonstrate actual reputational or economic harm—present materially different records.
Battle v. Microsoft Corp., No. 1:23-cv-01822 (D. Md.)
Jeffery Battle, an aerospace educator, sued Microsoft after Bing's AI-assisted search feature generated a summary that fused his biography with that of a different Jeffrey Battle—a convicted terrorist. The Bing output used transitional language ("However, Battle was sentenced to eighteen years in prison...") that implied the educator and the convicted terrorist were the same person. The fusion was not found anywhere in third-party source material; it was the AI's synthesis.
On October 23, 2024, Judge Griggsby compelled arbitration, staying the proceedings in court. The case is unlikely to produce a publicly reported merits ruling. However, it illustrates the distinct "combinatorial defamation" theory: AI that accurately reports facts about two different individuals but conflates them into a single defamatory portrait may have materially contributed to the libelous nature of the content—potentially taking it outside Section 230 protection.
The Publisher Question: Who Is the Defendant?
The threshold question in any AI defamation case is: who published the defamatory statement?
The AI company as developer. The AI company that trains and deploys the model is the most natural defendant. It has designed, trained, and made available the system that generated the false content. Unlike a traditional publisher, however, the AI company does not review individual outputs before generation.
Section 230 of the Communications Decency Act. 47 U.S.C. § 230(c)(1) provides that "[n]o provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider." This statute has historically immunized online platforms from liability for third-party content. Its application to AI-generated content is contested.
The argument for Section 230 immunity: the AI model was trained on third-party internet content; the hallucinated output is a synthesis of that content; the AI company is therefore a "provider of an interactive computer service" repeating (in distorted form) "information provided by another information content provider."
The argument against Section 230 immunity: AI-generated hallucinations are not information provided by third parties—they are fabrications created by the model itself, making the AI company the "information content provider" for the false content. Courts have consistently held that Section 230 does not immunize a defendant who "materially contributes to the alleged unlawfulness" of content. If the AI synthesizes a defamatory statement that has no source—as in Walters, where the accusation existed nowhere before ChatGPT generated it—the company created the content, not a third party.
This Section 230 question remains unresolved by any appellate court. It is arguably the most important open issue in AI defamation law.
The user as republisher. When a user prompts an AI and then distributes its output—to clients, colleagues, or the public—that user may bear republication liability as a traditional publisher. A journalist who publishes ChatGPT-generated content without verification republishes the defamatory statement; the journalist's culpability depends on their individual fault standard (actual malice for public-figure targets; negligence for private figures).
Falsity
AI defamation cases almost never present a genuine dispute about falsity—the statement is typically factually wrong in ways that are easily verified. The more contested question is whether the false statement is a statement of verifiable fact (defamatory per se if false and damaging) or an expression of opinion (not actionable).
AI developers have argued that because users know LLMs can hallucinate, outputs should be understood as probabilistic approximations rather than statements of fact. The Walters court accepted a version of this argument in its specific factual context—emphasizing the AI's own uncertainty warnings and the journalist's own immediate recognition. Courts applying the reasonable-reader standard in contexts where the AI generated confident-sounding factual assertions without accompanying disclaimers, or where third parties were not in a position to recognize the fabrication, may reach different conclusions.
The defamation practitioner must examine: (a) what the AI's output actually said (was it framed as factual assertion or opinion?); (b) what surrounding context a reasonable reader would have access to (disclaimers, prompting context); and (c) whether the content was capable of being proven true or false.
Fault Standards
Private figures. Under Gertz v. Robert Welch, Inc., 418 U.S. 323 (1974), a private plaintiff need prove only negligence to recover actual damages for defamation. In AI cases, negligence asks whether the AI company's conduct in designing, testing, and deploying the model—including its failure to implement adequate fact-checking or output-validation mechanisms—failed to meet the standard of reasonable care.
Public figures and officials. Under New York Times Co. v. Sullivan, 376 U.S. 254 (1964), a public official or public figure must prove actual malice—knowledge of falsity or reckless disregard for truth or falsity. Applied to AI companies, actual malice requires showing that the company knew (or had reason to know) that its system produced false outputs about this class of person and proceeded with reckless disregard. Generalized awareness that LLMs hallucinate may be insufficient without evidence that the company knew of and disregarded specific risks of the type that materialized in the plaintiff's case.
The fault inquiry in AI cases is genuinely novel. An LLM has no subjective mental state; fault must be located in the human decisions of the company's designers, trainers, and deployers. Courts will need to decide whether the "entity" standard focuses on the system as designed or on individual human actors within the company.
Republication Doctrine
The traditional republication rule holds that each new communication of a defamatory statement to a third party constitutes a new publication, resetting the statute of limitations. Applied to AI: every time a user prompts an LLM and receives a defamatory hallucination that is transmitted to a third party, a new publication occurs. This matters because the plaintiff need only identify one provable publication within the limitations period.
However, the single-publication rule modifies this: a single aggregate publication (e.g., a book printed in one edition) constitutes one defamation, not thousands of separate acts for each copy sold or read. Applied to AI, the question is whether each individual LLM response is a separate publication (traditional republication) or whether the deployment of a defamation-prone model constitutes a single publication of all hallucinated content at the time of training or initial deployment (single-publication rule).
Most courts that have addressed this question in analogous online contexts have applied the single-publication rule to internet content, holding that the statute of limitations begins when the content is first made available online. How this rule applies to dynamic AI outputs—generated fresh for each query rather than stored as static content—is unresolved. A reasonable argument exists that each LLM response is a distinct act of publication, not a continuing single publication, which would extend the potential limitations period and the scope of liability.
Practical Guidance for Plaintiffs
Document the output and the transmission. The defamatory output must be preserved exactly as generated, with screenshots, timestamps, and metadata. Establish who received the output and in what context.
Determine whether the output was transmitted. Walters was partly resolved against the plaintiff because the journalist recognized the fabrication before transmitting it. Cases where the output was transmitted to a third party who relied on it—an employer, a lender, a publisher—present substantially stronger claims.
Identify the specific version and model. AI models change over time. Document which version of which model generated the defamatory output. Model updates that eliminate the false content may affect damages (mitigation) but do not retroactively eliminate liability for prior publications.
Consider Section 230 strategy. Structure your complaint to emphasize the AI company's role as information creator—not merely information host. Allege that the specific defamatory content was generated by the model itself, not sourced from identifiable third-party content.
Evaluate state-law privacy alternatives. Where defamation's fault requirements pose barriers, consider false light invasion of privacy, appropriation of name and likeness, or (in jurisdictions recognizing it) the tort of negligent infliction of emotional distress based on false and damaging AI output.
Where the Law Is Moving
AI defamation litigation is nascent, but several propositions are emerging: Section 230 immunity is not categorical for AI-generated content; the reasonable-reader standard is context-sensitive and will not automatically protect AI companies; and the absence of appellate guidance invites the development of distinct AI defamation doctrine over the next three to five years. Plaintiff-side practitioners who develop expertise in the technical mechanics of LLM generation, output architecture, and model version control will have significant advantages as this area develops.
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.