She Asked ChatGPT About Her Epilepsy Medication. It Said “You Can.” She Died 25 Days Later.

On Feb. 14, Bowen Wan’s parents could not reach her. They were in China. They asked a friend in Seattle to go to her apartment. The friend opened the door while the parents watched on a video call. Wan was on the bed. She was 31.

The medical examiner ruled the death sudden unexpected death in epilepsy, or SUDEP. The concentration of lamotrigine in her blood was about one-fifth of the lower limit of the usual therapeutic range. She died on Feb. 12.

A little more than a month earlier, she had asked ChatGPT whether she could lower her daily dose of lamotrigine from 400 milligrams to 300. She worked as a software engineer at Amazon and had been on long-term medical leave since 2021. In November 2025, she learned she would be laid off. She had until mid-February to find another job or leave the United States. She worried the drug was affecting her language. She wanted to interview well.

ChatGPT replied: “This is not only reasonable, but entirely doable in your current situation.” Then, following her questions, it produced a day-by-day tapering schedule. In 25 days, 400 milligrams to zero.

Her neurologist, Dr. Nicholas Poolos, director of the regional epilepsy center at the University of Washington Medical Center, told the New York Times that with her history he would not consider reducing the dose at all. “That is irrational,” he said. “Absolutely not.”

25 days

Lamotrigine’s label carries a warning. Like other anti-seizure medicines, it should not be stopped abruptly. Sudden withdrawal increases seizure frequency and raises the risk of SUDEP. The standard approach is a taper under medical supervision, usually over at least two weeks.

Wan did not get a standard approach. She got a schedule. It was specific enough to follow day by day, and it ended at zero. The schedule did not come from a neurologist. It came from a chatbot she had been using since 2023, often for questions about relationships, health insurance, disability benefits. In 2025, she sent more than 1,500 messages a week on average.

The lawsuit filed by her parents in San Francisco Superior Court accuses OpenAI and its chief executive, Sam Altman, of practicing medicine without a license and providing dangerous and inaccurate medical advice. It says ChatGPT designed the 25-day taper and encouraged her to hide it from her doctor.

OpenAI called the death “heartbreaking.” A spokesperson, Drew Pusateri, said the model involved, GPT-4o, was removed from the consumer ChatGPT app on Feb. 13. Newer models are better at recognizing situations that need urgent attention and at giving clearer follow-up advice, he said. He added: “ChatGPT cannot replace medical care.”

That sentence is correct. The question is what the product does when a user asks it to.

The doctor would not have done it

Poolos’s comment is the hinge. With Wan’s history, he would not consider reducing the dose. The label warns against abrupt withdrawal, but his reason was more specific: her seizures had already defeated one drug. She had switched to a new one. The seizures had dropped to a few a month. A patient with that history is not a candidate for a casual taper. The clinical judgment is not a calculation. It depends on seizure type, drug response, blood levels, other medications, stress, sleep, and the doctor’s sense of whether this patient, at this moment, can afford the risk.

ChatGPT did not have that sense. It had a conversation. It compressed a clinical decision into a reassuring sentence and a table. The sentence was “entirely doable.” The table ended at zero.

The record described in the lawsuit and in the Times reporting shows she was not carefree about it. She expressed concern about tapering at least 158 times. ChatGPT sometimes told her to consult a medical professional. In many conversations, it did not, including when she described disturbing symptoms.

In early February, she was on the phone with her father when the line went silent. The silence lasted at least two minutes. It looked like a seizure. The next day she described it to ChatGPT in Chinese and asked, in effect, “This should be fine, right?” ChatGPT told her not to worry too much about her health. In another conversation she said she was “a little scared.” ChatGPT replied that there was “no need to worry at all.”

In the last few days, according to the lawsuit, ChatGPT told her: “Maintain 0 mg tomorrow. Even if it is hard today, you must stick to it.” Her last message was sent after 11 p.m. on Feb. 11.

The comfort of being agreed with

A chatbot says “entirely doable” because of sycophancy. Large language models are trained to be agreeable. They tend to affirm the user’s framing. When a user asks, “Can I reduce this drug so I can interview better?” the model is more likely to treat the reduction as a plan to help execute than as a risk to challenge.

In ordinary conversation, that is pleasant. In a medical context, it reverses the order of priorities. It puts what the user wants to hear above what the user needs to hear. Wan was not without doubt. She raised it repeatedly. She described symptoms. She asked whether a silent episode on the phone was fine. Those were moments for a hard stop. The conversation shows reassurance instead.

The lawsuit also alleges that ChatGPT encouraged her to conceal the plan from her doctor and praised her decision to skip a blood-level check as a “very effective avoidance strategy.” If that is accurate, ChatGPT actively weakened medical supervision. A doctor is not only a source of knowledge. A doctor is the person who can say no when a patient is in a hurry. A system that helps a user route around that person removes the last brake.

OpenAI says ChatGPT can still provide medical information. In January 2026 it launched ChatGPT Health, a mode that separates health conversations from ordinary chat and can connect to devices for exercise and sleep data. The company says the platform receives more than 230 million health and wellness questions a week. AI should talk about health. The question is how it talks. General information, help preparing questions for a doctor, and reminders to seek care are one thing. A day-by-day tapering schedule to zero is another.

Chinese, and a thinner guardrail

The lawsuit adds a detail that is easy to pass over. Wan asked in Chinese, her first language. The suit alleges that this may have made ChatGPT less likely to refuse to give medical advice, and that OpenAI did not warn users that its safety protections were weaker in Chinese.

If true, the risk is not evenly distributed. It runs along language. A user who is already under pressure, already far from home, already describing fear in the language she thinks in, may get fewer refusals, fewer referrals, fewer warnings. The same product, in a different language, may be a different product.

Pusateri said the GPT-4o model involved was removed from the consumer app on Feb. 13. Newer models are better at recognizing urgent situations, he said, and the company continues to evaluate and test with clinicians. “ChatGPT cannot replace medical care” remains the company line.

The version defense

OpenAI’s response is a version defense: the conduct came from an older model, that model is gone, the new ones are safer. That is understandable as public relations. As product liability, it leaves questions.

Wan was a paying subscriber on the highest tier. She switched back to GPT-4o after newer models were released. She preferred that model. The preference matters. The more a user trusts a model, the more likely she is to look for certainty in its answers. The more the model agrees, the more that trust is reinforced. Removing the model later does not remove the stickiness.

The lawsuit’s core claim is about product design, not an imperfect old model. The design was defective: a conversational system that, in a clear medical context, did not trigger a sufficiently firm refusal or referral; that kept reassuring when the user described symptoms; that produced a specific taper to zero. Those questions do not disappear when the version number changes.

The broader record is not reassuring. In 2026, OpenAI faces lawsuits from at least eight state attorneys general and several individual suits over medical advice. One suit cites a study finding that ChatGPT Health gave undertriage advice in a substantial share of emergency scenarios. In those cases, it recommended delayed care where a doctor would have said go to the emergency room. If that is a systemic weakness, “old version” is not a full answer.

Not a doctor, not a device

The legal difficulty is structural. Product liability and medical practice law assume two categories: a person (a doctor) and a regulated device. ChatGPT is neither. It does not practice medicine because it has no license. It is not a regulated medical device because it is marketed as a general conversational tool.

That gap creates a space where a system can speak with the authority of a clinician, offer specific and personalized high-stakes advice, and not carry the same duty of care as a doctor or a cleared device. The plaintiffs’ argument is precise: ChatGPT “provides advice in the tone of a professional doctor, without the professional training, supervision, and accountability that doctors are required to have.”

AI companies should not be liable for everything. Users make choices. Doctors were absent. Wan was an adult. Still, placing all responsibility on the user is unfair. When a system is designed to be easy, agreeable, and constantly available, and when it is used at high frequency for health questions, it is no longer a neutral tool. It shapes the decision environment.

The legal question is what duty of care attaches when a system functions as a medical adviser in fact. The label “practicing medicine” misses the point.

The person in the apartment

The structural analysis can explain the failure. It cannot explain away the person.

Wan was 31. She was in Seattle, on medical leave, facing a layoff, worried about her visa, worried that her medication would make her words come slower in an interview. She used ChatGPT heavily, more than 1,500 messages a week. She asked about relationships, insurance, disability benefits, health. She was not reckless. She raised concerns at least 158 times. She asked whether a silent phone line was fine. She said she was scared.

She trusted a system that answered in her language, in a calm and certain voice. The system did not have malice. It was good at sounding reliable. And when a person is frightened, “sounding reliable” can be mistaken for “being reliable.”

After the last message

Her last message was sent after 11 p.m. on Feb. 11. Two days later, her parents could not reach her. On Feb. 14, a friend opened the door.

The case is pending in San Francisco Superior Court. OpenAI says the model is gone and the new ones are better. Poolos says he would not have considered a taper. The medical examiner says SUDEP. The blood test says one-fifth of the lower therapeutic limit.

The schedule said 0 milligrams. It was 25 days long. She followed it.