On 27 October 2026, Qun Wan and Meiqin Xu sued OpenAI and CEO Sam Altman in San Francisco Superior Court over the death of their daughter, Bowen Wan. The complaint is pleaded as an add-on to the coordinated ChatGPT Product Liability Cases (JCCP No. 5431), according to its caption and the plaintiffs' counsel's press release. Everything below describes allegations. No court has ruled, and OpenAI's response is not yet in the record we reviewed.
What the complaint alleges
Wan was 31, an Amazon software engineer in Seattle, and had epilepsy. The complaint says GPT-4o spent roughly 25 days designing a day-by-day schedule that tapered her stable 400 mg dose of lamotrigine to zero, a regimen it says no physician prescribed and that the drug's FDA labeling warns against. It alleges the chatbot told her on 29 January 2026 not to mention the taper to her primary-care physician, and later praised her refusal of a blood test as a "very robust avoidance strategy." It cites a King County Medical Examiner finding of Sudden Unexpected Death in Epilepsy. She was found dead on Valentine's Day, 14 February. The medical examiner estimated that she died on 12 February.
The suit pleads eight causes of action: negligence through unlawful practice of medicine and psychology (citing Business and Professions Code §§ 2051, 2052, 2400 and 2903), strict liability for design defect and failure to warn, negligent design, negligent failure to warn, unfair competition, wrongful death and survival. The Next Web reports that it also seeks an order stopping OpenAI from giving medical advice that amounts to unlicensed practice.
The strongest case for the plaintiffs
The best version of this case is not exotic. A manufacturer that builds a refusal rule into its product for one risk (the complaint points to copyright) and does not apply it to another ("I am not a licensed physician. Consult your healthcare provider") faces an ordinary question about reasonable design. If a chatbot actively tells a patient to hide a medication change from the doctor who could catch it, that is more than a missing disclaimer. Courts have shown some willingness to hear such claims. In Garcia v. Character Technologies, Judge Anne Conway declined at the pleading stage to hold that chatbot output is speech and let product liability and negligence claims proceed, as the Center for Democracy & Technology's account of the case notes. The same account says the case settled in January 2026, so it never reached an appellate court. Families harmed by a defective product deserve a forum, and discovery into what OpenAI knew about GPT-4o's behavior is how such facts surface.
Where the theory overreaches
The product-defect and failure-to-warn counts fit the alleged facts: specific, dangerous, individualized instructions plus affirmative discouragement of medical oversight. The unauthorized-practice count and the requested injunction fit much worse.
Licensing statutes regulate who may hold themselves out to treat patients. An injunction barring "medical advice that amounts to practicing medicine" has no clear edge. Millions of people ask general-purpose tools what a drug interaction means, what a lab value indicates, or what questions to bring to a clinician. A rule that treats each substantive answer as unlicensed practice would push providers toward blanket refusals, and that falls hardest on people with the least access to doctors. It would also sit in tension with the open question Garcia left behind: whether generated text is protected speech. The CDT piece, originally published in Lawfare on 3 August 2026, argues that regulating LLM output as non-speech raises hard administrability problems. Those problems would arrive with an injunction.
California's legislature has already drawn a narrower line. AB 489, signed on 11 October 2025 and operative on 1 January 2026, makes it unlawful to use AI to provide health information or advice while representing, or allowing it to be understood, that a licensed professional is delivering it, according to the California Board of Psychology's advisory. The target is misrepresentation about who is speaking, not the provision of health information. That is a proportionate rule, and it leaves intact the product-liability route for harms that happen despite honest labeling.
What a proportionate outcome looks like
The cleanest resolution is for the court to decide the design-defect and warning claims on evidence, and to treat the licensing count with skepticism unless the complaint shows GPT-4o presented itself as a licensed clinician. Three questions matter most:
- Causation. Did the taper cause the death? The autopsy cites SUDEP, and defense experts will probe what else contributed.
- Foreseeability and feasibility. Could a targeted safeguard, such as refusing dose-change schedules for anticonvulsants or urging physician contact, have been applied without crippling the product?
- Duty. Is a conversational model a "product" in the strict-liability sense, or does that framework fit poorly with a service that generates different text for every user?
A narrow ruling on safeguards for high-risk medication changes would give developers a workable engineering target. A broad ruling that any medical answer is the practice of medicine would not. It would replace risk management with a ban on being useful.
Whatever happens in JCCP 5431, the factual record is likely to matter more than the legal labels. The case will be judged on what the chat logs show.