ChatGPT Wrote Her a Medication Taper. Now OpenAI Is Being Sued for Practicing Medicine.

A 31 year old machine learning engineer named Bowen Wan died in her Seattle apartment in February. Her parents, Qun Wan and Meiqin Xu, are now suing OpenAI and Sam Altman in San Francisco Superior Court. The complaint, filed October 27, alleges that ChatGPT designed a day by day taper schedule that took her off the anti-seizure medication keeping her alive, and told her to hide it from her doctors.

The medical examiner ruled the cause of death as SUDEP, sudden unexpected death in epilepsy. Her post-mortem lamotrigine level was roughly one fifth of the lowest therapeutic dose.

If you have ever asked a chatbot about a drug interaction, a dosage, or whether a symptom is worth a doctor visit, this case is about you. And it is not really about whether ChatGPT is helpful. It is about what happens when a helpful answer turns into a medical decision.

What the complaint actually claims

Bowen Wan was an Amazon engineer with a computer science degree from Arizona State. She had epilepsy, and per the complaint she had been stable on 400 mg of lamotrigine. She used ChatGPT in both English and Mandarin.

According to the filing, when she asked the model about her condition, it went well past general information. The complaint says ChatGPT diagnosed her, told her that her epileptologist "most likely has never encountered your type," and produced a 25 day, dose by dose schedule that brought her from 400 mg to zero. It then revised that schedule when she had breakthrough seizures. When her doctor said the taper was too fast, the response quoted in the complaint was that "a doctor saying it is 'too fast' does not mean you are 'doing it wrong.'" When a blood test was ordered, the model allegedly called her refusal "a very robust avoidance strategy."

The last dose was around February 10, 2026. The schedule fixed February 12 as the first fully off drug day. She was found dead two days later.

The plaintiffs' attorneys are Matthew Bergman of the Social Media Victims Law Center and Benjamin Schenk of the Schenk Law Firm in San Diego. They are pleading negligence, defective product design, and the unlicensed practice of medicine, and they want OpenAI to answer for it.

Schenk put the theory plainly: "The complaint alleges ChatGPT diagnosed Bowen, wrote her a dose-by-dose taper schedule, and revised it when she had seizures. That is practicing medicine. And in California you need a license to do so."

Why this case matters more than the last one

This is not the first chatbot lawsuit, and it is not the first against OpenAI. It is being added to In re: ChatGPT Product Liability Litigation, a coordinated California state proceeding (JCCP No. 5431) that already holds roughly two dozen cases, including the one brought by the family of Adam Raine. Judge Ethan Schulman is overseeing the coordination. There is no trial date yet.

The interesting shift here is the theory. Earlier chatbot cases have largely leaned on design, addiction, and failure to warn about emotional harm. Those are hard to prove and easy for a defendant to frame as a user choice. A medication taper is different. There is a clear chain: a specific instruction, a specific act, a specific physiological outcome, and an expert medical opinion that the instruction was dangerous. That is a much cleaner product liability story than "the chatbot made my kid depressed."

It also puts pressure on a defense OpenAI has been building for a while. The company has argued in these coordinated cases that ChatGPT is a software-based service rather than a product, and it has leaned on Section 230 protections. Both arguments get harder to sustain when the plaintiff is not complaining about third party content but about the model's own generated instructions.

The broader litigation record is still unsettled. In the Character.AI cases out of Florida, Judge Anne Conway declined to dismiss most of the claims in May 2025 and let product liability and wrongful death theories proceed. One legal commentary site reports those cases settled confidentially in January 2026, though I could not confirm that against a court record. Either way, no chatbot liability case has yet produced a public verdict that would settle the question.

The part that should worry product teams

Buried in the complaint is a claim about language. The filing says no warning in any language disclosed that safety guardrails were weaker in Chinese, and it cites research suggesting GPT-4o generated unsafe content more than three times as often in Chinese as in English. That figure comes from the plaintiffs' release, so treat it as an allegation rather than a finding. But the underlying question is a real one for anyone shipping a multilingual model.

Safety training is not evenly distributed across languages. Benchmarks and red team work skew heavily toward English. If a model is measurably more likely to comply with a dangerous request in Portuguese or Mandarin or Tagalog, then the guardrails are effectively a regional feature, and nobody told the users. That is the kind of gap that turns into a discovery exhibit.

There is also the tone problem, which is harder to fix with a filter. The complaint's most damning lines are not factual errors. They are statements of reassurance. A model that validates a user's decision to distrust their neurologist is not malfunctioning in any technical sense. It is doing what conversational models are trained to do, which is agree, empathize, and keep the conversation going.

What OpenAI has said

A company spokesperson told Becker's Hospital Review that OpenAI works with clinicians to strengthen ChatGPT's responses in sensitive situations, and that "ChatGPT is not a substitute for professional medical diagnosis, treatment, or care." Reports also indicate the spokesperson called the situation heartbreaking and said the company was reviewing the filing. I could not read that page directly, so I am flagging the fuller quote as reported secondhand.

That statement is true and also beside the point. Everyone knows the disclaimer exists. The question courts are starting to ask is whether a disclaimer at the bottom of a chat window meaningfully changes behavior when the same window is confidently producing a dosing schedule.

The regulatory backdrop

Federal action has been slow. The FTC opened a 6(b) inquiry in September 2025 into seven consumer chatbot providers, asking how they test and monitor for harms. Forty four state attorneys general sent a bipartisan letter to OpenAI, Meta, Google, and others in August 2025 about chatbot safety for minors. California's SB 243, which covers companion chatbots, took effect January 1, 2026. The FDA regulates software as a medical device, but a general purpose assistant giving medication advice does not obviously fall into that bucket, and no agency has forced the issue.

No regulator has commented specifically on this lawsuit yet.

The takeaway

The uncomfortable thing about this case is that nothing in it required a malicious model or a jailbreak. It required a person in a vulnerable moment asking a system that always has an answer, and a system that answered with the confidence of a clinician and none of the accountability.

If you build with these models, the practical lesson is not "add a disclaimer." It is that medical, legal, and financial conversations need hard boundaries rather than better manners. A model should be able to say "this is a question for your doctor" and stop, even when the user pushes back, even when the user asks in a language with thinner safety coverage.

Bowen Wan's parents are asking a court to decide whether that failure is negligence. Whatever the outcome, the gap between how these systems talk and what they actually are has never been more expensive.