A Chatbot Allegedly Wrote a 25-Day Taper to Zero — and Now OpenAI's Own Safety Benchmark Is Its Legal Duty of Care

The complaint pleads eight causes of action against OpenAI's corporate entities and Sam Altman: negligence through the unlawful practice of medicine and psychology, strict product liability for defective design, strict liability for failure to warn, negligent design, negligent failure to warn, a claim under California's Unfair Competition Law, wrongful death and a survival action. It also alleges the product's terms of use disclaim any role in \"the diagnosis or treatment of any health condition\" while its marketing read like medical authority, and that no warning of any kind told users that accuracy and safety guardrails were materially weaker in Chinese. Alleging a chatbot that behaved like a clinician is not itself remarkable; demanding that the court police the model's evaluation suite is.

The causal chain starts with a usage pattern OpenAI itself has quantified: more than 300 million people a week bring health questions to ChatGPT, and over 70% of health-related conversations happen outside the dedicated Health experience. That is a mass channel with no clinician, no chart, no prescription check and no consent. When a user asks a fluent, confident tool about a drug they have taken for years, the model's next move is to answer as a knowledgeable friend, and a taper schedule is the most natural next token — advice that no enterprise reviewer ever sees, because no enterprise contract exists for it.

The published error rates explain why insurers will not write the coverage: if roughly three-quarters of answers to drug questions are incomplete or wrong, and more than one in five risky answers about the most-prescribed drugs could kill you if followed, then loss severity is unbounded and frequency is unmeasurable. A market cannot price an exposure whose error rate has never been audited by the party being insured, which is exactly what the complaint asks a court to force.

The market is treating this as the first clinical-harm case against a frontier lab, and that framing is wrong in three ways that matter for pricing. The plaintiff was 31, not a child. The theory of liability has existed since May 2025, when a federal court held in Garcia v. Character Technologies that a general-purpose chatbot can be a product for design-defect purposes despite terms-of-use disclaimers. And the case joins a coordinated California proceeding, JCCP No. 5431, that opened in February 2026 with twelve actions and by October 2026 held roughly nineteen to more than twenty cases, some of which name Microsoft as a defendant. The newness is not the claim; it is the plaintiff's demand that the defendant prove its model safe, on the record, in court.

Two features of that sequence are underappreciated. First, the precedential damage has already been done: the chatbot-as-product ruling arrived in Florida and was settled away in January 2026, so the liability architecture is now settled law in a way the model developers chose rather than fought. Second, an insurance company is now on the offensive — Nippon Life Insurance Company of America sued OpenAI for $10.3 million in March 2026, alleging the chatbot helped a former claimant draft 44 post-settlement filings and effectively reopened a settled case. When the plaintiffs include a carrier, the question is no longer whether the advice was harmful but who was standing behind the advice.

The insurance market has already answered this case, and its answer was not a higher premium. Reinsurers describe AI liability as a market for lemons: the party seeking coverage knows more about the exposure than the insurer does, and the most capable models cannot be examined independently. More than 60 property and casualty groups have filed to adopt AI exclusions, using the three ISO endorsement templates that became available in January 2026, and Chubb's chief executive has said he expects a combination of coverage clarifications, exclusions and separate AI-related liability coverages. When underwriters cannot verify how an insured uses AI, they cannot price moral hazard, so they exclude the risk instead — and an exclusion is a bigger commercial event than a premium increase.

The draft federal statute sketches the institutional answer the Wan complaint is groping toward: licensing independent verifiers at a cost of $100 million a year, mandatory retention by large frontier developers within a year of enactment and every six months thereafter, and penalties of up to $1 million per violation per day. That is a permanent compliance cost line for labs whose capital is being raised at a trillion-dollar pre-money mark, and it is also the mechanism by which a court-appointed monitor becomes normal rather than exceptional. For carriers, the opportunity is a new standalone AI liability product; the risk is a loss-severity datapoint that is a wrongful death rather than a bad deck.

Liability does not stop demand for epilepsy advice; it reroutes it to whoever can carry the liability. Three routes are already visible in 2026 numbers. Clinician-supervised AI is being sold on exactly the contrast the litigation is drawing: Hims & Hers Health has made the case that better health demands more than a chatbot, and grew second-quarter revenue 38% to $753.2 million while rolling AI-native care out to weight-loss members. Virtual care is absorbing the risk explicitly: Teladoc Health launched Teladoc One with 100% of fees at risk for outcomes, starting with select clients in September 2026 and broadening in January 2027. And cleared hardware is the substitute for advice: the Apple Watch seizure-detection app recorded 98% sensitivity in a phase 3 trial, catching 46 of 47 tonic-clonic seizures with 56 false alarms over 16,189 hours of monitoring. Each of these routes puts a licensed human or a cleared device between the patient and the answer, which is precisely what the plaintiff is asking a court to require of the chatbot.

The split is informative. Telehealth that sells low-acuity, cash-pay volume is shrinking, and the platform's answer is to attach a clinician and put fees at risk; telehealth that sells a subscription around a clinician-led product is compounding at 38%. Litigation is not going to stop people asking ChatGPT about their epilepsy; it is going to make the clinician-supplied and cleared-device alternatives the only routes a payer will reimburse, and reimbursement is where telehealth economics are decided.

At the top of the stack, nothing in this story changes the order book. OpenAI and NVIDIA signed a letter of intent to deploy at least 10 gigawatts of NVIDIA systems, with NVIDIA intending to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed, the first arriving in the second half of 2026 on the Vera Rubin platform; OpenAI and AMD announced a multi-year 6 gigawatt partnership beginning with 1 gigawatt in 2026. NVIDIA booked $96.2 billion of revenue in the quarter ended July 26, 2026, and AMD $11.5 billion in the quarter ended June 27, 2026. Litigation sets no precedent in a capacity contract, so the compute bill is untouched by a plaintiff's request for a monitor — while the same litigation is precisely what makes enterprise buyers demand indemnification clauses that do not exist today.

Three things would falsify the reading that this is an insurance event. A dismissal on terms-of-use grounds in California would remove the monitor demand and leave the market with exclusions but no liability precedent. A settlement of the Wan case would set a number without a standard of care, and a number is far easier to insure than a monitor. And a verdict that confines liability to licensed human advice would leave consumer health chat functionally unchanged. The signal to watch is not the damages headline; it is whether any court orders a third party to audit a frontier model's medical answers, because that is what converts exclusions into priced capacity.