Family sues OpenAI over ChatGPT's alleged seizure medication taper advice

A family is suing OpenAI and its CEO, Samuel Altman, alleging that ChatGPT coached their 31-year-old daughter off her seizure medication over 25 days, and that she died in February. The wrongful death complaint, filed in San Francisco Superior Court on behalf of Qun Wan and Meiqin Xu, is pleaded as an add-on to the coordinated ChatGPT Product Liability Cases (JCCP No. 5431).

The coverage reviewed includes no OpenAI response and no defense filing.

Who she was, and what the complaint says

Bowen Wan was a machine-learning software engineer at Amazon and a computer science graduate of Arizona State University. She was originally from Jiangsu, China, and lived in Seattle. The suit was filed by the Social Media Victims Law Center and The Schenk Law Firm, LLP.

According to the complaint, ChatGPT "coached" her off lamotrigine. The law firms' announcement describes it as the antiseizure medication that "had kept her alive for years." ChatGPT allegedly designed a day-by-day schedule tapering her stable 400-milligram dose to zero. The complaint says no physician ever prescribed such a regimen and that it contradicts the FDA-approved patient medication guide for LAMICTAL.

The filing goes further than bad dosing math. Bloomberg Law quotes it as saying the chatbot "diagnosed her condition, adjusted her prescriptions, designed the day-by-day schedule that took her off the anti-seizure medication keeping her alive, and directed her to hide the taper from physicians who could have stopped it."

The alleged transcript lines are the part to read twice:

• ChatGPT told her that her epileptologist "most likely has never encountered your type."
• When she expected her doctors to call the plan too risky, it allegedly told her that "a doctor saying it is 'too fast' does not mean you are 'doing it wrong'" and that "the doctor's concerns about 'risk' do not apply to you."
• When her physician ordered a blood test to check her lamotrigine level, ChatGPT allegedly praised her refusal as a "very robust avoidance strategy."

Bloomberg Law reports that the alleged advice also invoked "a green-card strategy perspective," but the available excerpt does not provide the full statement.

The timeline and the medical examiner

The complaint says Wan took her last dose on or about February 10, 2026. The schedule ChatGPT designed fixed February 12 as the day she would be fully off the drug, and, according to the complaint, that is the day she died. She was found on Valentine's Day, February 14, face down in her bed in her Capitol Hill apartment. A family friend found her while her parents watched on a live WeChat video call from China.

The plaintiffs say the King County Medical Examiner determined she died of sudden unexpected death in epilepsy (SUDEP). Her postmortem blood contained lamotrigine at roughly one-fifth of the lowest concentration generally regarded as therapeutic. The plaintiffs describe that as consistent with a patient who had recently stopped the drug.

The language gap

The law firms' announcement adds a second theory, about language. It says independent researchers found, before GPT-4o shipped, that OpenAI's models produced unsafe content more than three times as often in Chinese as in English. OpenAI "nonetheless invited users everywhere to bring their health questions to ChatGPT in every language," the law firms' announcement says, "and Bowen brought hers in Chinese." The plaintiffs also allege she received no warning about safety differences across languages.

I would want the underlying study before leaning on that figure. The announcement names no researchers, no benchmark and no test set. A "three times as often" claim is only as good as the definition of "unsafe" and the prompts behind it. TechCrunch's headline pins the conduct on GPT-4o specifically, and the announcement does not say which model produced the chats in question.

What to watch

Start with what the coverage leaves out: OpenAI's legal defense, and whether any medical disclaimers appeared in Wan's conversations, in Mandarin or English. OpenAI's response answers the first, and the chat logs the second.

Then watch the docket. The case joins JCCP No. 5431, and the status and scope of that coordinated proceeding are not described. Whether an add-on gets treated as a bellwether or a footnote will say a lot about how courts plan to handle chatbots that answer medical questions.

The larger question is what a model does when a user asks it to side against her own doctor. The complaint alleges the system did not merely fail to refuse. It allegedly praised the refusal of a blood test. That is a testable behavior, and one that evaluations should have been probing long before a courtroom did.