For two years the central question in AI injury litigation has been a question of category. Is a large language model a product, whose outputs are the product working as designed, or is it a service, whose outputs are speech that Section 230 and the First Amendment place beyond the reach of an injured user? The plaintiff bar has argued product. The defense bar has argued service. The fight has been doctrinal, slow, and unresolved.
On the evening of June 12, 2026, the federal government walked into that fight from an entirely different door, and without meaning to, it answered the question in the plaintiff's favor.
A takedown, not a recall
The facts are narrow and worth stating precisely, because the temptation to overread them is the first trap. Anthropic released Claude Fable 5 and its restricted sibling, Mythos 5, on June 9. Three days later the Commerce Department sent Anthropic's chief executive a letter placing both models under export controls, barring access by any foreign national whether inside or outside the United States, including the company's own foreign-national employees. Unable to verify citizenship on every request in real time, Anthropic disabled both models worldwide within hours. The models were live for roughly seventy-two hours. By the company's account, the directive supplied no specific national-security rationale.
This was not a product recall. It was not a tort judgment. It was not a finding that any model was defective or that any output had injured anyone. It was an export-control measure, run through the deemed-export framework, the same legal machinery that once regulated strong encryption as a munition during the crypto wars and that governs advanced semiconductors today. A tort court adjudicating whether AI output is a product is bound by none of it. Anyone who draws a straight line from "the government shut it down" to "therefore it is a product" has made a category error a first-year associate could dismantle.
The value of the episode is not precedential. It is in the characterization the government chose, and could have avoided.
The characterization the government chose
Export controls regulate things and capabilities. They do not regulate speech, and for good reason: an attempt to suspend a model worldwide because of the ideas it might express would collide immediately with the First Amendment. The government did not frame Fable 5 that way. It framed the model as a controlled capability, a system whose capacity to generate certain outputs was dangerous enough, and inherent enough to the design, to justify pulling it from the market overnight.
Read what that framing assumes. It locates the danger in what the system is built to do, not in what any particular user asked it to say. A government does not impose deemed-export controls on a library because a patron might check out a dangerous book. It imposes them on the capability itself. To treat Fable 5 as a controlled capability is to treat it as a thing that does something, which is to say, a product. The executive branch reached past the speech frame and the service frame and picked up the instrumentality frame, the one the plaintiff bar has been arguing for since the first of these cases was filed.
What Garcia settled, and what it did not
The courts were already moving the same direction before Commerce sent its letter. In Garcia v. Character Technologies, the United States District Court for the Middle District of Florida held that Character A.I. was a product for purposes of claims arising from defects in the app rather than from the ideas or expressions within it. 785 F. Supp. 3d 1157, 1180 (M.D. Fla. 2025). The same court declined to hold, at the pleading stage, that the chatbot's outputs were speech protected by the First Amendment. Strict product claims, negligence, and a wrongful-death count survived dismissal.
The limits matter as much as the holding. Character.AI settled Garcia and several related suits in early 2026, before any appellate court reviewed the product ruling and before trial. So the order persuades; it does not bind. It is one trial court's treatment of one motion in one case that resolved before it could be tested on appeal. A piece that frames Garcia as settled precedent overstates it, and a careful reader will notice. The honest framing is that Garcia is the most developed judicial articulation of the product theory to date, that it survives as persuasive authority, and that the doctrinal question remains formally open. The Fable 5 directive points the same way Garcia points. Two signals, neither dispositive, both aimed at the same conclusion: output is design, design is product, product is liability.
Section 230 and the material-contribution problem
The characterization does its hardest work against Section 230, which is where most of these cases will actually be won or lost. The statute immunizes an interactive computer service from liability for information provided by another information content provider. The entire shield depends on a single premise: that the model's output is someone else's content. Knock out that premise and the immunity goes with it.
The statute supplies the tool to knock it out. Section 230(f)(3) defines an information content provider as any entity responsible, in whole or in part, for the creation or development of information. The Ninth Circuit, sitting en banc in Fair Housing Council v. Roommates.com, held that a platform forfeits immunity when it materially contributes to the content at issue, when it is responsible, even partially, for developing what is published. 521 F.3d 1157 (9th Cir. 2008). Partial responsibility for development is enough to strip the protection.
Apply that to a generative model honestly. The model does not store and serve a user's words back to a reader the way a message board hosts a post. It composes new text, often the only place that specific output has ever existed, in response to a prompt. The prompt is an input; the output is something the system created and developed. Under the material-contribution standard, a system responsible in whole or in part for developing the content is an information content provider, not a neutral conduit, and Section 230 does not reach it. The export directive sharpens the intuition by sheer force of authority. When a sovereign suspends a model because of what it generates, it is not treating the model as a passive host of third-party speech. It is treating the model as the source.
One caveat belongs in the argument, not buried beneath it. No court has yet squarely applied the material-contribution standard to large-language-model output, and the question of whether a model is legally "responsible for the creation or development" of what it generates is genuinely unresolved. The argument above is the strongest plaintiff-side framing of an open question, not a statement of settled law. Saying so makes the case harder to attack, not weaker.
The defendant who asked to be regulated
The most useful piece of material in the whole episode did not come from the government. It came from Anthropic's own chief executive, one day before the takedown. In an essay published the day after Fable 5 launched, Dario Amodei called for mandatory third-party testing of frontier models and for governments to be empowered to block systems that fail safety audits, modeling the regime explicitly on the Federal Aviation Administration. Models above a compute threshold, in his proposal, would face required audits in cybersecurity, biological-weapons risk, loss of control, and automated AI research. He rejected the idea that public concern about these systems is a marketing problem and called the concern accurate.
Consider what frame that adopts. The FAA analogy is a product analogy. The airplane is the textbook regulated product, certified for safety before deployment and subject to design-defect and failure-to-warn liability after it. When the most prominent figure in the defendant class reaches for "test it and certify it like an aircraft before release," he is characterizing the model as a manufactured product whose safety is a function of its design, not as a service and not as protected expression. The foreseeability value is just as direct. A manufacturer publicly affirming that the risks are real, and naming the categories, is notice-and-foreseeability material for a design-defect or warning theory.
The caveats here are real and a responsible piece names them. A call for prospective regulation is policy advocacy, not a concession that any specific model is defective or that any output is a product in the tort sense. Treat the essay as characterization and foreseeability, not as a party admission, because framing it as "the CEO admitted the product is dangerous" is a mischaracterization that weakens everything around it. The FAA frame also cuts the other way over the longer run: the certification regime Amodei wants is pre-market federal review, and if such a regime is ever enacted, defendants gain a compliance defense and a federal-preemption argument against state tort claims. The same essay that helps characterization today could hand the defense a shield tomorrow.
One further wrinkle will be raised against any piece that argues defendants treat these systems as products rather than speech. In its separate dispute with the Pentagon, Anthropic framed its refusal of the government's terms as an exercise of its First Amendment rights. A defense reader will wave that around. The distinction holds and should be stated plainly: that First Amendment claim concerned Anthropic's corporate decision to decline certain uses of its model, not a claim that the model's outputs are themselves protected speech. The two are not the same, but the point is worth preempting before someone else makes it.
Why defendants keep dodging
The strategic behavior of the defense bar is itself evidence of where the risk lies. In Walters v. OpenAI, when the model fabricated a false and defamatory claim about a radio host, OpenAI had Section 230 available and chose not to invoke it, winning instead on conventional defamation standards by arguing the output could not reasonably be read as a statement of fact. The reason to leave the shield untouched is obvious: a ruling that tested Section 230 against generative output could collapse the protection for everyone. In the consolidated AI proceedings, OpenAI's own filings characterize ChatGPT as a software-based service rather than a product. The defense is working hard, and visibly, to keep these systems classified as service and speech.
Anthropic's defense of Fable 5 fits the same pattern, and contains its own double edge. The company argued that the flagged capability is narrow, routine for security work, and widely available from other frontier models, OpenAI's included. As a response to an export order that singled out one company, that is a fair complaint. As a posture for product litigation it is also the industry-custom, state-of-the-art defense in embryo: a model that performs like its peers, the argument runs, is not defective. Plaintiff's counsel should expect that argument and meet it on its own terms, because availability is not safety, and an industry that uniformly ships the same risk has not thereby made the risk reasonable.
What actually reprices it
Set the rhetoric aside and the practical question is unchanged. Nothing that happened in June decided a case. The repricing event for the AI injury docket is still the set of dispositive-motion rulings expected across the wrongful-death cases in the second half of 2026, the moment a court tests the product theory and the Section 230 defense against a real record rather than a pleading. Those rulings will move inventory. An export directive and a CEO's policy essay will not.
What the June episode changed is the terrain on which those motions will be argued. For two years the defense has worked to keep these systems classified as service and speech, the output someone else's, the model a mere conduit. Garcia cracked the product question open at the pleading stage. The Fable 5 directive did something the plaintiff bar could not do for itself: it put the United States government on record treating a frontier model's output as a capability inherent in the product, dangerous by design, in the same week the model's own maker asked to be regulated like an aircraft manufacturer. The defense will keep arguing conduit. It is now arguing it against the executive branch and against the other side's own chief executive at once.
The whole product-liability theory sits in that contradiction, compressed to a sentence. Dangerous by design, and everyone ships it anyway.
The cases and statements described in this article are matters of public record and, where they involve pending litigation, reflect allegations and arguments that have not been finally adjudicated. Case status is current as of publication. The consumer-facing tracker of the AI injury docket at Lawsuit Informer is updated as the dockets move.