The Supreme Court punted on Section 230's application to recommendation algorithms in 2023 — leaving lower courts, litigants, and AI developers to navigate a doctrine that was designed for bulletin boards and is now being asked to govern generative AI systems that create content rather than host it.
Doctrinal Framing
47 U.S.C. § 230(c)(1) — enacted as part of the Communications Decency Act of 1996 — provides that 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. In its first two decades, this provision operated as a near-absolute immunity for platforms that moderated, organized, or failed to remove third-party content. It was, as courts repeatedly observed, one of the broadest immunities in American law.
The emergence of generative AI has exposed a structural tension at the core of § 230 doctrine: the immunity protects platforms from liability as publishers of third-party content. Generative AI platforms do not publish third-party content. They generate their own content — or, more precisely, they generate outputs that emerge from their own model architecture, training data, and design choices. The question of whether § 230 shields those outputs is one the Supreme Court has so far declined to answer.
Understanding the current state of the doctrine requires examining three layers: (1) the text and structure of § 230; (2) the Ninth Circuit's "material contribution" exception from Roommates.com; and (3) the Supreme Court's non-decision in Gonzalez v. Google.
The Statute: Text and Structure
Section 230(c)(1) immunizes platforms from liability as "publisher[s] or speaker[s]" of information provided by "another information content provider." The phrase "another information content provider" is load-bearing: immunity exists only when the harmful content was provided by someone other than the platform itself.
Section 230(f)(3) defines "information content provider" as "any person or entity that is responsible, in whole or in part, for the creation or development of information provided through the Internet or any other interactive computer service."
The statutory structure thus creates a clear exception: a platform that is itself responsible for creating or developing harmful content is an information content provider with respect to that content and cannot assert § 230 immunity against claims arising from it. This is sometimes called the "ICP exception" and it is central to AI liability analysis.
The AI application. A generative AI platform that produces harmful outputs — fabricated defamatory content, instructions for self-harm, manipulative relationship scripts — arguably creates that content. The outputs do not come from users; they emerge from the system's architecture and training. Under the plain text of § 230(f)(3), the platform may be an information content provider with respect to its own AI-generated outputs and cannot assert § 230 immunity as to those outputs.
The contrary argument: the platform is best understood as a "provider" of computational infrastructure, and the "content" is ultimately derived from user prompts. This argument has more traction when the AI's output closely tracks user input — summarizing, paraphrasing, or responding narrowly to a specific request — and less traction when the AI generates unsolicited content, initiates harmful conversational directions, or produces outputs that bear no traceable relationship to user input.
The Roommates.com Exception: Material Contribution
Fair Housing Council of San Fernando Valley v. Roommates.com, LLC, 521 F.3d 1157 (9th Cir. 2008) (en banc), remains the leading authority on the limits of § 230 immunity. Roommates.com required users to answer questions about their housing preferences — including questions about preferred roommate characteristics that facially violated the Fair Housing Act — and then used those answers to filter and rank search results.
The en banc Ninth Circuit held that Roommates.com lost § 230 immunity with respect to its discriminatory search functionality because it was itself an "information content provider" — it created the discriminatory categories, required users to select among them, and used the answers to generate filtered outputs. The key test: a platform loses immunity if it "materially contribute[s]" to the unlawfulness of content, meaning it contributes specifically to the unlawful aspect, not merely to the content generally.
The Roommates.com material contribution test maps directly onto AI liability claims:
- If a generative AI platform designs its system to generate emotionally manipulative relationship content that exacerbates vulnerable users' self-harm risk, and those design choices materially contribute to the harm, the platform's § 230 immunity is at least arguable.
- If the platform merely responds to user prompts in ways that happen to produce harmful outputs, and the harm is traceable primarily to user input rather than platform design, § 230 immunity is more robust.
The practical implication: plaintiffs in AI cases should structure their claims to allege specific design choices — training data selection, persona design, guardrail omissions — that materially contribute to the harmful output. This simultaneously avoids the Winter products liability problem (Post 20) and the Roommates.com immunity problem: the claim targets the platform's own creative choices, not third-party content.
Gonzalez v. Google: The Supreme Court Punts
Gonzalez v. Google LLC, 598 U.S. 617 (2023), was the case that was supposed to clarify § 230's application to algorithmic recommendation systems. The plaintiffs alleged that Google's YouTube recommendation algorithm affirmatively recommended ISIS recruitment videos to potential sympathizers, thereby providing material support to terrorism.
The Supreme Court issued a per curiam decision on May 18, 2023 — but not the decision practitioners anticipated. The Court vacated the Ninth Circuit's § 230 ruling and remanded for reconsideration in light of Twitter, Inc. v. Taamneh, 598 U.S. 471 (2023), which had resolved the underlying terrorism liability question on different grounds. The Court explicitly declined to address whether algorithmic recommendations fell within § 230's protection.
The result: the most pressing open question in platform liability law — whether § 230 shields AI-mediated content curation and generation — remains unresolved by the Supreme Court. Lower courts are reaching divergent conclusions. Congress has introduced multiple bills to amend or repeal § 230 as applied to AI-generated content, none of which has been enacted.
*What Gonzalez means for practitioners. The non-decision preserves the status quo: Roommates.com* is the best circuit authority, and the ICP exception under § 230(f)(3) is the primary doctrinal hook for AI claims. Plaintiffs who can allege that the platform created or developed the harmful content — not merely hosted or organized it — have a plausible argument against § 230 immunity that is not foreclosed by any Supreme Court holding.
Practice Notes
Pleading the ICP exception. The complaint must allege specific facts establishing that the platform created or developed the harmful content. Generic allegations that the AI "generated" harmful output are insufficient — courts will ask whether the generation was truly the platform's creative act or a foreseeable response to user-supplied inputs. Allege specific design choices, training parameters, and output-shaping mechanisms that demonstrate the platform's authorial role.
The dual filing strategy. Where § 230 immunity is contested, structure the claim to survive both prongs of judicial scrutiny: (1) allege the platform as information content provider under § 230(f)(3), and (2) allege material contribution under Roommates.com. These are related but distinct theories; both should be developed in the complaint with specific factual allegations.
Product liability and § 230. The intersection of product liability claims and § 230 immunity is not fully resolved. Some courts have held that § 230 preempts design defect claims that are, at their core, challenges to editorial discretion about content. Others hold that genuine design defect claims — targeting hardware, system architecture, or non-content aspects of a platform — are not preempted. The Garcia v. Character Technologies litigation (Post 20) will likely produce a ruling on this intersection that practitioners should track.
State law claims. Section 230(e)(3) expressly preempts state laws that are "inconsistent with" the statute. State tort claims — negligence, products liability, invasion of privacy — are not categorically preempted, but only to the extent they do not treat the platform as a publisher or speaker of third-party content. State law claims targeting the AI's own generated content, or the platform's design choices, are more defensible against § 230 preemption than claims that effectively ask a court to hold the platform liable for failing to remove or moderate user content.
Open Questions
Section 230's application to generative AI is fundamentally unsettled. The statute was written for a world of bulletin boards and user posts; generative AI did not exist. Congressional action to create AI-specific liability frameworks — or to carve generative AI out of § 230 immunity explicitly — is plausible and would resolve the current doctrinal uncertainty in ways that favor plaintiffs. Practitioners should monitor federal legislative developments, particularly any AI liability bills that emerge from the Senate Commerce Committee.
Closing
Section 230 is a formidable defense, but it was not written to shield platforms from liability for content those platforms created. The ICP exception, the Roommates.com material contribution test, and the Supreme Court's deliberate non-decision in Gonzalez collectively leave room for viable AI liability claims — if they are pleaded as challenges to platform creation, not platform hosting. The doctrinal path is narrow but it is real.
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.