What the Social Media Addiction Cases Foreshadow for AI Tort Litigation

As I previously discussed, Florida’s lawsuit against OpenAI and Sam Altman may help usher in the next major wave of litigation against technology companies. The 100-plus copyright cases pending against OpenAI, Anthropic, Meta, and others are the first major test of how courts will respond to large language models. But those cases may eventually be eclipsed by actions that resemble Florida’s product-liability and failure-to-warn theories.

The ongoing social media addiction litigation is not a perfect template for this nascent AI tort litigation, but it is an important preview. Future plaintiffs suing AI companies are likely to frame the alleged harms as flowing from product design and representations about safety rather than from third-party content.

Recent reporting confirms that AI companies are already focused on “what happens with this product liability approach,” including “the possibility of multiplying suits by state attorneys general.” The recent California state-court bellwether trial on social media addiction offers useful guidance, subject to two caveats. First, the distinctive nature of LLMs may lead courts to reach different conclusions than they have in cases involving social media platforms. Second, although social media plaintiffs have had early success, the legal landscape will remain unsettled until appellate courts address difficult questions involving federal statutory and constitutional law.

The Social Media Bellwether

Against that backdrop, the latest major development is the June 9, 2026 order from a California Superior Court judge denying Meta and YouTube’s motions for judgment notwithstanding the verdict and for a new trial after a jury found both companies liable for harm to a minor plaintiff and awarded millions in damages. The post-trial motions principally argued that Section 230 of the Communications Decency Act, 47 U.S.C. § 230, and the First Amendment barred the plaintiff’s theories of liability.

On Section 230, the court held that the statute did not warrant disturbing the verdict because “[t]here was substantial evidence that plaintiff was harmed by the design features of Instagram, regardless of any of the content found on the platform.” The court also emphasized that it had repeatedly instructed the jury “not to base liability on those activities for which Section 230 provided protection." Notably, the court signaled frustration with Meta's insistence on relitigating an argument the court had already "rejected several times" before trial, while observing that YouTube had "very reasonably" preserved its Section 230 arguments for appeal without belaboring them.

The court took a similar view of the First Amendment defense. The critical distinction, in the court's analysis, was between liability for content and liability for product design. The court distinguished Meta’s authorities, including Moody v. NetChoice, LLC, the Supreme Court’s 2024 decision recognizing First Amendment protection for social media companies’ editorial judgments and content-moderation decisions. Moody, the court explained, did not “address the very different question of whether liability for harm flowing, not from the content on the platform, but from the design features of the platform themselves, is precluded by the First Amendment.”

Given Meta’s resources and the stakes these cases pose for Instagram and its other products, the company may litigate the issue through appeal and potentially to the Supreme Court. After the court denied its post-trial motion, Meta issued a blunt statement previewing that strategy: “The plaintiffs’ legal theory attempts to improperly circumvent Section 230 and the First Amendment, and we expect this ruling to be overturned on appeal.”

Section 230's Diminished Role in AI Cases

Even if Meta or another social media company ultimately prevails on Section 230 or First Amendment grounds, those defenses are likely to be less formidable in litigation involving large language models. Section 230 generally protects tech companies from being treated as the publisher or speaker of information provided by a third party. That framing fits social media more naturally because the disputed content is often a post, image, or video supplied by a user. It fits generative AI less comfortably because the allegedly harmful output is generated by the model itself, rather than merely supplied by another user. To be sure, some defendants may argue that LLM outputs are derived from or synthesized from third-party training data, and thus still involve "information provided by another information content provider" under Section 230. But that argument is difficult to sustain where the model does not reproduce any particular third-party content and instead generates novel text in response to a user prompt. While social media companies can argue that they are being sued over harmful third-party content, an argument that has carried the day many times before, that argument is considerably less straightforward in the AI context. Indeed, in the handful of individual tort suits brought against AI companies to date, Section 230 has not played a central role.

Two public statements from 2023, still early in the generative-AI era, illustrate the challenge for AI companies. During the Senate Judiciary Subcommittee’s May 16, 2023 hearing on “Oversight of A.I.: Rules for Artificial Intelligence,” OpenAI CEO Sam Altman was asked about Section 230 during an exchange with Senator Lindsey Graham and appeared to distance generative AI companies from broad reliance on that protection. And during the 2023 oral argument in Gonzalez v. Google, Justice Gorsuch suggested that generative-AI outputs may not be covered by Section 230.

The First Amendment as the Next Battleground

Recognizing the limits of Section 230, AI companies have instead emphasized that the First Amendment limits liability for harms allegedly caused by model outputs. If Section 230 is an imperfect fit for generative AI, the First Amendment is likely to become the more important defense. But that argument faces threshold difficulties. Most fundamentally, it remains unsettled whether generative AI outputs should be treated as “speech.” Even if they are, courts will still need to distinguish claims targeting expressive output from claims targeting model design, safety, and the absence of reasonable guardrails. On one hand, LLM outputs are mere statistical predictions, i.e., mathematical operations that select the most probable next token based on patterns in training data, and thus lack the communicative intent that has traditionally anchored First Amendment protection. But others posit that the outputs are functionally indistinguishable from human-authored text and serve the same informational and expressive purposes that the First Amendment is designed to protect. 

Failure-to-Warn Theories

Failure-to-warn theories are also likely to become more important in AI tort litigation than they have been in the social media cases. Plaintiffs can be expected to argue that AI companies knew or should have known about foreseeable risks from chatbot interactions, including emotional dependency, hallucinations, reinforcement of delusional or self-destructive thinking, use by minors or vulnerable users, generation of dangerous instructions, and exposure of sensitive personal information, yet failed to provide adequate warnings or safety interventions. Those theories will not eliminate First Amendment questions, but they offer courts a familiar product-liability framework without treating every claim as a direct attack on expression.

Key Takeaways

That is why the social media addiction cases are worth watching closely. They foreshadow a litigation strategy built around product design rather than content, and they may provide early appellate guidance on where Section 230 and the First Amendment end and ordinary tort principles begin. But AI cases will not track the social media cases perfectly. LLMs do not merely host or recommend third-party material. Rather, they generate responses, simulate conversation, adapt to users, and are increasingly used in high-stakes settings, including military and national security contexts. Those differences may make AI product-liability litigation harder for defendants to dismiss at the outset — and harder for courts to fit within existing internet-law doctrines.

For companies developing or deploying AI systems, the lesson is clear: the time to evaluate product-design choices, warning frameworks, and potential liability exposure is now, before the litigation landscape crystallizes and before plaintiffs' theories are tested at scale. The social media cases demonstrate how quickly once-novel claims can mature into existential litigation.