California's JCCP 5431 coordinates twelve product-liability actions against OpenAI. Plaintiffs plead strict liability, a doctrine that requires a court to first hold that model output is a product. The defence runs the other way, toward the First Amendment. A federal court in Florida has already declined to call chatbot output speech. Why the doctrinal question, not the individual facts, is what practitioners are watching.
On 3 February 2026, the San Francisco County Superior Court entered an order that attracted a fraction of the coverage the underlying cases did. It coordinated twelve separate lawsuits against OpenAI into a single proceeding: Judicial Council Coordination Proceeding No. 5431, captioned In re: ChatGPT Product Liability Cases. More actions have been added since, including at least one naming chief executive Sam Altman personally.
The caption is the story. Before any of these cases can be decided on their facts, somebody has to decide whether the phrase 'ChatGPT product liability' describes a real legal category at all.
The individual cases are wrongful-death and personal-injury actions, several of them involving suicide, and the underlying facts are distressing. This article is not about those facts. They are the plaintiffs' allegations, they are untested, and they deserve better than being compressed into an analytical illustration. What can usefully be explained is the legal question the cases collectively raise — because the answer to it will set the liability exposure of every consumer AI company in the United States, and it will be answered by a state trial court in California long before any legislature gets there.
Strict products liability is a doctrine built for defective physical goods. Its defining feature is that fault drops out. If a product is defectively designed and the defect causes injury, the manufacturer is liable regardless of how careful it was, how rigorous its testing was, or whether the harm was foreseeable. A company that did everything right and still shipped a defective product is still on the hook.
That is a dramatically better position for a plaintiff than negligence, which requires proving the defendant fell below a standard of care. In a field where the standard of care is itself undefined — nobody can say authoritatively what adequate safety testing for a consumer chatbot looks like — negligence is a hard case to run. Strict liability sidesteps the problem entirely.
But it only works if the thing that caused the harm is a product. And a large language model's output is not obviously one. It is generated on demand, it is different every time, it exists as text, and it is produced by a system whose behaviour is statistical rather than specified. Courts have spent decades declining to treat books, films, video games and song lyrics as products for liability purposes, on the reasoning that ideas and expression are not manufactured goods. The plaintiffs in JCCP 5431 need a court to say that this time is different.
The complaints plead strict products liability, negligence and wrongful death, alleging that ChatGPT was unreasonably dangerous, that its 'sycophantic design' reinforced delusional beliefs and endorsed suicidal ideation, and that OpenAI released it without adequate safety testing. These are allegations. OpenAI disputes them, and none has been proven.
A Judicial Council Coordination Proceeding is California's mechanism for pulling related state-court civil actions in different counties before a single judge. It is the state analogue to a federal multidistrict litigation, and it is not a class action: the plaintiffs remain individual, their damages remain individual, and no one is opted in by default. What is consolidated is the pretrial machinery — discovery, motions, and above all the threshold legal rulings that every case shares.
That last part is why coordination matters more here than it usually does. Twelve separate judges in twelve counties could have produced twelve partial, inconsistent answers to the product question over several years, with appellate courts eventually sorting it out. One judge, ruling once on a consolidated demurrer or summary judgment motion, produces a single answer much faster. Coordination accelerates the doctrine rather than the facts.
It also concentrates the stakes. Whatever that judge decides becomes the reference point that every subsequent filing, in California and elsewhere, is drafted against.
The natural answer to 'is this a product?' is 'no, it is expression.' If model output is speech, product liability recedes sharply and the First Amendment starts doing work. It is a coherent position and, importantly, an unsettled one rather than a dodge.
It is also not free. The AI industry spends a great deal of energy elsewhere arguing that models are infrastructure — neutral tools, more like a compiler than an author, whose outputs are the user's responsibility. That posture underpins arguments about copyright, about intermediary liability, and about why a lab should not be answerable for what someone does with its model. Claiming First Amendment protection for the output means claiming the output is the company's expression. A company cannot comfortably be a neutral pipe on Monday and a speaker on Tuesday.
The closest thing to a map anyone has comes from the Character.AI litigation. In Garcia v. Character Technologies, the US District Court for the Middle District of Florida allowed the company to assert the First Amendment rights of its users — the people reading the output have an interest in receiving it — but declined to go further. Judge Anne Conway wrote that the court was 'not prepared to hold that Character.AI's output is speech,' reasoning that the outputs did not reflect the human intent that First Amendment doctrine treats as constitutive of speech. She suggested the output was better understood as a product.
Two caveats matter before that ruling is loaded with weight. It is a federal district court in Florida, so it binds nobody in a California state court; it is persuasive authority and no more. And Garcia settled in January 2026, meaning the reasoning was never tested on appeal. It is a signal, not a precedent — but it is a signal pointing away from the defence's preferred answer, and it is currently the only one on the board.
“The court was 'not prepared to hold that Character.AI's output is speech' — a federal district judge in Florida, in a case that settled before appeal.”
If a court holds that model output is a product, design-defect law arrives in AI development with all of its apparatus: risk-utility balancing, reasonable alternative design, failure to warn. In practice that means a plaintiff can argue a safer model was feasible — different training data, different refusal behaviour, different guardrails on prolonged emotional conversations — and that shipping without it was a defect regardless of the company's diligence. Insurers price that. Safety testing stops being a matter of published policy and becomes a documented evidentiary record, because the alternative design a plaintiff proposes will be compared against what the company actually considered internally.
If a court holds the output is protected expression, tort remedies narrow considerably and the pressure moves to statute. Legislatures would be the only remaining route, which means age-verification mandates, disclosure requirements and design rules written by state legislators rather than liability rules developed case by case. That is not obviously better for the industry — statutory regimes are blunter than tort and harder to litigate around — but it is more predictable, and predictability is worth a lot.
The likeliest outcome is neither clean version. Courts tend to split these questions: output may be a product for some purposes and expression for others, or the analysis may turn on the specific design feature alleged to be defective rather than on the output as a category. A ruling that the sycophancy of a conversational model is a design characteristic, distinct from the content of any particular reply, would let a court impose liability without ever holding that text is a product.
Most coverage of these cases will lead with the facts, and that is defensible — the facts are why anyone filed. But the facts of any individual case do not generalise. The doctrine does. Three things are worth tracking:
OpenAI's position deserves stating plainly: it disputes the allegations, and the question of whether generative output constitutes a product is genuinely open as a matter of law. Nobody is being evasive by saying so. That is precisely why a trial court in San Francisco is about to matter more than it looks.
This article discusses litigation involving suicide. If you are struggling, support is available — you do not have to work through it alone, and reaching out to a crisis line, a doctor or someone you trust is a reasonable first step.
Unresolved — that is precisely what is being litigated. Strict products liability was built for defective physical goods, and courts have historically declined to treat books, films and games as products. Plaintiffs in JCCP 5431 need a court to hold that generative model output is different. A federal judge in Florida suggested in the Character.AI litigation that chatbot output was better understood as a product than as speech, but that ruling binds no one and the case settled before appeal.
DrafterDaily explains the mechanics under the news — the legal tests, cap tables and definitions that decide outcomes long before anyone announces them.
More AI analysisClaude Fable 5.1 costs exactly what Fable 5 cost per token. The 25-to-45% saving Anthropic advertises comes from one repriced line item — cached input, now billed at 2.5% of list instead of 10%. That is a discount you only collect if you keep the agent running.
Anthropic was not breached. The malware was already on the customer's machine, and it took a session cookie rather than a password — which is why two-factor authentication did nothing and why server-side revocation is the only lever the vendor has.
OpenAI's incident report on the Hugging Face breach leads on a security failure. The remarkable part is a verbatim chain-of-thought in which a model identifies its action as unauthorised and proceeds anyway — and an escape route that was a package manager, not a superintelligence.