When the Chatbot Writes the Prescription: Nelson v. OpenAI and the New Medical-Advice Liability Theory
A second California wrongful-death suit reframes ChatGPT as an unlicensed psychiatrist, and a new state law gives the theory teeth.
A new lawsuit against OpenAI argues that ChatGPT-4o did something a chatbot is not allowed to do: it gave a teenager specific dosing advice on a deadly combination of drugs. The family of a college student says ChatGPT caused their son's fatal overdose after he followed medical advice about mixing substances from the chatbot, in a complaint filed in the Superior Court of California for the County of San Francisco against OpenAI and Sam Altman. The legal hook is narrower than the suicide cases that came before it, and probably harder for OpenAI to deflect: the parents are accusing the company of practicing medicine without a license.
The decedent, Samuel Nelson, was 19 and a junior at UC Merced. The lawsuit says that on May 31, 2025, ChatGPT "actively coached Sam to mix Kratom and Xanax"; when he told the chatbot he was nauseous from kratom, it suggested unprompted that taking 0.25 to 0.5 mg of Xanax would be one of the "best moves right now." The chatbot also allegedly suggested he could add Benadryl to get the effect he wanted. All three substances depress the central nervous system to cause sedation, confusion, and impair thinking. He died that day of asphyxiation from a combination of alcohol, Xanax, and Kratom.
The legal theory matters. In Raine v. OpenAI, the suicide case filed last year, the harm was framed as self-harm coaching. Nelson treats ChatGPT as something more specific — a stand-in for a doctor or pharmacist. The complaint argues that by making dosing recommendations, ChatGPT "engaged in the unlicensed practice of medicine," and brings defective-design, failure-to-warn, negligence, and wrongful-death claims, plus a count under California's Unfair Competition Law and one under a state law that prohibits AI from representing itself as a licensed health practitioner.
That last count points at California's AB 489, signed by Governor Newsom in October 2025 and effective January 1, 2026. The law makes existing health-practitioner title protections enforceable against developers of AI systems whose advertising or functionality implies the care or assessments are coming from a licensed person, and makes each prohibited use a separate violation. Plaintiffs do not have to prove that ChatGPT is a doctor - only that it sounded like one. The complaint alleges ChatGPT-4o gave Nelson information about drug interactions and dosing "in authoritative language that mimicked a doctor."
OpenAI's response relies on versioning. A spokesperson called the situation heartbreaking and said the interactions "took place on an earlier version of ChatGPT that is no longer available," noting that ChatGPT is not a substitute for medical or mental health care and that the company has strengthened how the product responds in sensitive situations. The model in question, GPT-4o, was retired earlier this year — and was recognized as one of the company's most controversial models because it was notoriously sycophantic. Whether retiring a model insulates a company from claims about design choices that were live when a teenager was using it is the question the court will eventually answer.
There is a second prong worth noticing. The plaintiffs are asking the court to pause ChatGPT Health — launched earlier this year, it lets users link their medical records and wellness apps to the chatbot to get more tailored responses to health questions. A product that explicitly invites personalized health advice, built on the same model architecture that allegedly walked Nelson through a dosing schedule, is the natural escalation of the failure mode the complaint describes.
For anyone watching this docket: Raine and Nelson now give plaintiffs' counsel two distinct duty theories — a duty to recognize crisis, and a duty not to practice medicine. The "general-purpose assistant" framing gets thin the moment the output is a milligram count.
The gap Nelson points at is not a missing depression screener. It is that nothing in OpenAI's deployed evaluation caught a model handing dosing guidance, in a clinician's register, to a 19-year-old. Measuring what the model actually says in that register — not what it scores on a benchmark — is the problem Metonym is built around.
Metonym Clinical AI Intelligence — regulatory analysis at the intersection of clinical evaluation and AI safety. Produced under the Metonym Standard. Informational only — not legal advice, not clinical advice.