Section 230 in the Age of Generative AI

Few statutes of the early internet era have proven as durable, or as controversial, as Section 230 of the Communications Decency Act of 1996. Its operative clause is famously brief: no 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.¹ For more than a quarter century that single sentence has supplied the legal scaffolding upon which the social web was built — sheltering platforms from defamation suits over user posts, from negligence claims premised on hosting decisions, and from a long catalogue of state-law theories that would otherwise have made the moderation of user-generated content commercially impossible. The Fourth Circuit's foundational reading in Zeran v. America Online² set the interpretive tone, and the lower courts have, with relatively few exceptions, followed it for a generation.

The arrival of large language models complicates this settled architecture in ways that the 1996 Congress could not have anticipated. The statute's central distinction — between the platform that hosts third-party content and the "information content provider" who authors it — was drawn against the background of a clear empirical division of labor. Users typed messages; platforms relayed them. When a generative model embedded within a platform produces an answer to a user's prompt, the analytic clarity of that division begins to dissolve. The question is no longer whether the platform transmitted what a user wrote, but whether the platform itself, through its model, became the speaker.

The narrow holdings of Gonzalez and Taamneh

The Supreme Court had a recent opportunity to clarify the contours of platform immunity in Gonzalez v. Google LLC³ and its companion case Twitter, Inc. v. Taamneh.⁴ The plaintiffs in Gonzalez sought to hold YouTube liable under the Anti-Terrorism Act for serving algorithmically recommended videos produced by the Islamic State; Taamneh presented an analogous theory against Twitter and other platforms. The Court resolved the cases on a narrow basis. In Taamneh, it held that the plaintiffs had failed to plead aiding-and-abetting liability under the underlying federal statute, and in Gonzalez it accordingly vacated and remanded without reaching the Section 230 question at all.

For observers awaiting an authoritative pronouncement on whether algorithmic recommendation constitutes the platform's own speech for Section 230 purposes, the result was a doctrinal non-answer. The Court was plainly aware of the significance of the question and equally plainly disinclined to resolve it on the record presented. What emerged was an implicit invitation to the lower courts and Congress to continue developing the doctrine through more finely calibrated cases — an invitation that has not gone unanswered.

The question is no longer whether the platform transmitted what a user wrote, but whether the platform itself, through its model, became the speaker.

The "material contribution" line and its strain

The most useful precedent for analyzing AI-generated content is not Gonzalez but the Ninth Circuit's en banc decision in Fair Housing Council v. Roommates.com.⁵ There the court held that a platform loses Section 230's protection when it "materially contributes" to the unlawfulness of the content at issue — in that case, by designing a dropdown menu that compelled users to express discriminatory housing preferences. The conceptual move in Roommates was to treat the platform's design choices as a form of authorship when those choices structurally produced the unlawful content rather than merely conveying it.

Generative AI puts pressure on this framework from both directions. On one hand, a model's output is contingent on a user's prompt, and one might argue that the user remains the proximate "information content provider" — much as a search engine's results are shaped by, though not authored by, the user's query. On the other hand, the model produces, with no meaningful intermediation, novel text that did not previously exist anywhere. To call the prompter the author of the output, as one would call a person the author of a Google search result, is to obscure the magnitude of the model's own contribution. The output is substantively a product of the model's parameters, training data, and inference procedure; the prompt is merely the occasion for its generation.

Early appellate signals and the defamation problem

Lower courts are beginning to confront these questions in concrete cases. The defamation claim asserted against OpenAI in Walters v. OpenAI⁶ — in which a Georgia radio host alleged that ChatGPT had fabricated false statements attributing financial misconduct to him — was resolved on summary judgment in the defendant's favor on actual malice and fault grounds, but the court's analysis assumed without deciding that the model's output could, in principle, be attributed to the platform for purposes of state defamation law. Other district courts have begun to follow suit, generally declining to extend Section 230's protections to model-generated text on the theory that the platform is the relevant "information content provider" when its own software authored the statement.

The implications are substantial. If model outputs are not third-party content, they fall outside Section 230 entirely and are governed by the ordinary common law of defamation, intentional infliction of emotional distress, negligence, and the various state-law consumer-protection statutes. The fault standards of New York Times v. Sullivan and Gertz v. Robert Welch, Inc. remain available defenses, but the existential question — whether the platform may be sued at all — would be answered in the affirmative.

What a thoughtful doctrinal response might look like

It is tempting to declare that Section 230 is simply obsolete for the generative era and to await congressional intervention. That declaration would be premature. The statute has always required a careful inquiry into whether the content at issue was "provided by another information content provider," and the existing doctrine — including the Roommates.com material-contribution test — is adaptable, in principle, to the new technological setting. What the doctrine cannot do is paper over the fact that, when a model generates a sentence, the platform's role has shifted from conduit to author. The First Amendment may yet protect much of what such a model produces, and the common law of defamation has its own demanding fault requirements. But the categorical immunity that Section 230 was understood to confer upon the transmission of third-party speech does not translate, without substantial doctrinal revision, to a setting in which the platform itself is the speaker.

The Review will continue to track the development of this doctrine in the appellate courts. For related discussion of platform liability and the constitutional treatment of algorithmic decision-making, see our commentary on algorithmic content curation and on the constitutional limits of state platform regulation.