Large Libel Models: Liability for AI Output with UCLA School of Law Professor Eugene Volokh

Episode Summary

Executive Summary: Eugene Volokh explains how existing defamation law may apply to AI-generated falsehoods, arguing that AI companies can be liable when their systems themselves output defamatory claims rather than merely linking to third-party content. He distinguishes public-figure and private-figure standards, discusses Section 230 limits, and suggests companies may need stronger verification and filtering to reduce liability as AI becomes more widely used.

Main Topics: Defamation law applied to AI output (Priority: 5/5): Volokh’s article examines whether AI-generated false statements about real people can support libel claims against AI vendors such as Microsoft or OpenAI. Context and whether AI output is treated as factual (Priority: 5/5): He explains that whether output is presented as fiction, opinion, rumor, or factual search results affects whether defamation law applies. Damages and standards for public vs. private figures (Priority: 5/5): The discussion covers different liability thresholds and available remedies depending on whether the subject is a public figure, private figure, or involved in a matter of public or private concern. Section 230 and AI-created content (Priority: 5/5): Volokh argues Section 230 protects companies when they republish third-party content, but not when their own systems generate the defamatory statement. User reliance, amplification, and downstream responsibility (Priority: 4/5): The hosts and Volokh discuss whether users who receive or act on AI output can be liable, and when their reliance becomes negligence or malpractice. Future regulation and broader AI risks (Priority: 4/5): The episode ends with speculation about whether courts, Congress, or industry practice will shape AI liability, and how military and existential risks complicate regulation.

Key Arguments: AI companies may face defamation liability when their systems themselves synthesize and output false claims about identifiable people, because the content is attributable to the company rather than a third party. Context matters: a clearly fictional prompt or obvious joke is less likely to be defamatory, while a search-like or answer-like presentation is more likely to be taken as factual. Warnings such as “there might be errors” or “rumor has it” do not automatically defeat a libel claim if the statement is still conveyed as a factual allegation. For public figures, liability generally requires proof of knowledge of falsity or reckless disregard; if a company is alerted to a false output and fails to correct it, that may strengthen the claim. For private figures, negligence may be enough in some cases, especially if the falsehood causes actual harm; in matters of private concern, damages can be broader. Section 230 likely does not shield companies from liability for outputs their own systems generate, because the statute covers information provided by another content provider, not the defendant’s own speech. Users who merely see inaccurate AI output are generally not liable, but professionals who rely on it without checking may face malpractice or competence problems. Current AI systems are useful precisely because they generate plausible synthesized text, but that same feature creates legal risk when the output is false and reputationally damaging. The law may adapt through ordinary application of existing defamation principles, though Congress could choose to intervene if a workable policy solution emerges. AI progress could lead to major changes in legal process and governance, but the panel notes uncertainty about whether the technology will plateau or accelerate into broader use.

Data Points: Date of episode: Thursday, July 27th - Opening introduction to the podcast episode Professor Volokh’s law review articles: Over 90 - Host’s introduction describing his scholarship Years as a computer programmer: 12 years - Host’s introduction to Volokh’s background Age at graduation: 15 - Host notes Volokh graduated in science/math/computer science at age 15 Supreme Court clerkship: 1 clerkship - Volokh clerked for Justice Sandra Day O’Connor Example damages threshold for public figures: Knowledge or recklessness required - Volokh explains the First Amendment standard for public figures/public officials Example damages threshold for private figures: Negligence may suffice - Volokh explains private-figure defamation standards Potential AI quote verification rule: Never output quotation marks unless verified - Volokh suggests a minimum safeguard for AI systems Hypothetical arbitration speed/cost: 30 minutes for $3,000 - Volokh describes a possible AI arbitrator scenario compared with litigation Hypothetical litigation alternative: 3 years and $300,000 - Used to contrast traditional legal processes with AI arbitration Company examples: Microsoft, OpenAI, Bing, ChatGPT - Named as defendants or systems relevant to AI defamation claims Notable case names: Jeffrey Battle; Walters v. OpenAI - Examples discussed as live defamation disputes involving AI output

Pivotal Quotes: "Can you sue Microsoft? Can you sue OpenAI for letting out into the wild this software that produces what appears to be defamatory material?" — Eugene Volokh: He frames the central legal issue: whether AI vendors can be liable for harmful generated output. "The key difference is that Bing itself... is outputting. It is information that is provided by the defendant himself." — Eugene Volokh: Explaining why Section 230 may not protect AI-generated defamatory statements. "I think chat GPT and Bing especially, a search engine output, I think people use because they think it's going to be pretty reliable much of the time." — Eugene Volokh: He argues that AI search-style outputs are taken seriously and can cause reputational harm even with error warnings.

Implications: AI companies may need stronger verification, filtering, and response systems to avoid defamation exposure. Courts will likely adapt old libel doctrine first, but major design and policy changes may follow as AI becomes more trusted and widely used.