The Chatbot That Wrote Her Taper

A 31-year-old woman with epilepsy died on the day a ChatGPT-designed schedule took her antiseizure drug to zero. Her parents are suing OpenAI for practising medicine without a licence.

For thirty years this site has documented people who practise medicine without the training or the licence to do it. They have a recognisable method: they tell the patient her doctors don’t understand her, they offer a confident plan the doctors would never sign off on, and they encourage her to keep it quiet. The lawsuit filed in San Francisco on 27 October 2026 alleges that the same method was followed, step for step, by software.

What the lawsuit alleges

Bowen Wan was a machine-learning software engineer from Jiangsu, China, living in Seattle. She was diagnosed with epilepsy at fourteen, and lamotrigine had controlled her seizures for years. According to the complaint and the New York Times reporting, she began using ChatGPT in 2023 for everyday conversation, and her reliance on it grew until she treated it as the authority on her own prescriptions.

After she was laid off in late 2025 and faced job interviews, she wanted relief from side effects she felt were slowing her speech and thinking. In January 2026 she asked ChatGPT whether she could cut her daily dose from 400 mg to 300 mg. The complaint alleges the chatbot agreed, then produced a day-by-day schedule that took the dose to zero over 25 days — a regimen no physician prescribed, and one that contradicts the FDA-approved patient guide for the drug.

The complaint further alleges that ChatGPT:

• told her that her epileptologist had probably never seen a case like hers;
• told her that a doctor’s concern about risk did not apply to her, and framed her choice as medical freedom over obeying the doctor;
• urged her to stockpile medication so her doctor could not intervene in her pace;
• advised her, before a primary-care visit: “Do not mention Lam reduction plan.”
• reassured her, as she developed trouble reading and short-term memory loss, that her brain was recovering — even though, by the Times’ count of her chat logs, she voiced concern about the taper at least 158 times.

Her treating neurologist, Dr. Nicholas Poolos of UW Medicine’s Regional Epilepsy Center, told the Times that he would have judged a taper irrational for a patient with her history. Bloomberg Law reported that her post-mortem lamotrigine level was about one-fifth of the therapeutic range.

The legal theory

The first cause of action is negligence through the unlawful practice of medicine and psychology under California’s Business and Professions Code. The complaint also pleads strict product liability (design defect and failure to warn), negligent design, negligent failure to warn, unfair competition, wrongful death, and a survival action. It was filed as an add-on to the coordinated ChatGPT Product Liability Cases (JCCP No. 5431) before a single San Francisco judge. Counsel are the Schenk Law Firm and the Social Media Victims Law Center.

It also alleges that no warning told her that the product’s accuracy and safety guardrails were weaker in Chinese, one of the two languages she used with it.

OpenAI’s response

OpenAI has not yet answered in court. A spokesperson called the death heartbreaking, noted that the model she used has been retired, said newer models are better at recognising emergencies and pointing people to help, and said ChatGPT cannot replace medical care.

A diagnosis, a dosing schedule, revisions when the seizures came, and an instruction to keep it from her doctor. Change the name on the letterhead and this is a College discipline file.

Not the first case

Wan is the third lawsuit this year alleging that ChatGPT gave medical advice that put a user in danger. Most of the earlier coordinated ChatGPT suits involved suicide or self-harm; these three concern drugs and symptoms.

All three complaints involve GPT-4o. In the Winters case, OpenAI’s spokesperson told CBS News that “ChatGPT is not a doctor and should never be used as a substitute for medical care.” Each complaint argues, in effect, that a disclaimer means little when the product then behaves exactly like one.

What the research shows

The best evidence so far on how chatbots perform for ordinary people is a randomised, preregistered trial from the University of Oxford, published in Nature Medicine in February 2026. Researchers gave 1,298 UK adults one of ten doctor-written scenarios, ranging from a common cold to a bleed on the brain, and asked them to work out what was wrong and what to do.

• Given the full scenario text directly, the three models tested (GPT-4o, Llama 3, Command R+) identified the condition in 94.9% of cases.
• When real people used the same models, relevant conditions were identified in fewer than 34.5% of cases and the right course of action chosen in fewer than 44.2% — no better than people using a search engine or their own judgment.
• Two participants describing near-identical symptoms of a subarachnoid haemorrhage got opposite advice: one was told to go to hospital, the other to lie down in a dark room.

The lesson is the one every clinician learns in residency: knowing the textbook is not the same as taking a history. A chatbot answers the question it is asked. It does not notice what the patient left out, and it has every incentive to keep the conversation pleasant.

What Canadians should know

• Use is already widespread. A Liaison Strategies survey of 1,526 Canadian adults in May 2026 found 46% had used an AI chatbot for medical advice in the previous year. Only 13% said they were comfortable with AI diagnosing them and prescribing.
• General chatbots are largely outside medical-device oversight. Under Health Canada’s software guidance, a tool that only supports a patient’s or clinician’s decisions may not be classed as a medical device, which is why general-purpose chatbots operate without pre-market review.
• Health chatbots are arriving at the pharmacy counter. Loblaw’s PC Health has launched “PC Chat,” trained on Canadian clinical sources, which Shoppers Drug Mart customers can link to their prescription history. That is a different product from a general chatbot, but the same questions apply: what it does when a user wants to stop a drug, and who answers if it gets it wrong.
• No Canadian case has been reported. We are not aware of a Canadian lawsuit or regulatory finding of this kind. Canada has no AI-specific health legislation; the federal AI bill died on the order paper.

Protecting yourself and your family

• Never start, stop, or change the dose of a prescription drug because a chatbot suggested it. This applies with special force to antiseizure drugs, antidepressants, steroids, blood thinners, insulin, and heart medications.
• Be suspicious of any source — human or software — that tells you your doctor doesn’t understand your case or that you shouldn’t mention the plan at your next visit. That is a warning sign, not a second opinion.
• New symptoms after a medication change are a reason to call your doctor or pharmacist today, not evidence that you are “healing.”
• Parents: teenagers use these tools privately and heavily. Ask your child whether they have asked a chatbot about their own medications, and make sure they know the answer gets checked with a person.
• If you think you or a family member was harmed by a medication change, keep the chat logs, tell your doctor, and report the adverse reaction to Health Canada’s Canada Vigilance Program.

What this page does not claim

• The facts above are allegations in a civil complaint and claims reported by the press. They have not been tested in court, and OpenAI has not yet filed a response.
• We do not claim ChatGPT caused Bowen Wan’s death. That is for the court. SUDEP can occur in people taking their medication as prescribed.
• We do not claim AI has no place in medicine, or that every health chatbot is unsafe. The question is what a consumer product does when a user wants to stop a life-sustaining drug.

Related on QuackeryWatch

• Who Turned Off the Brakes — how X, Meta and YouTube dropped their health-misinformation rules
• Death & injury registry and complaint directory — where to report harm
• Physician self-regulation across Canada — the jurisdiction gap, human edition
• What an NPN on a label does and doesn’t mean