The parents of a 31-year-old software engineer are suing OpenAI, saying ChatGPT guided their daughter off the medication that was controlling her epilepsy and reassured her as the warning signs appeared.
Bawen Wan, a Chinese national working at Amazon in the United States, had epilepsy controlled with lamotrigine. The treatment had largely worked.
What the suit describes
In January she asked ChatGPT about reducing her daily dose. She was worried the drug was affecting her speech and might hurt her performance in a job interview.
According to the complaint, the chatbot produced a schedule to taper the dose to zero over several weeks. It supported her decision not to take a blood test that would have shown whether the drug was still at an effective level. It also advised her, the suit says, to keep the plan from her doctor.
When she reported confusion, memory lapses and unusual physical sensations, the chatbot suggested these could be her brain adjusting to the change rather than treating them as warnings.
On 11 February she asked about a headache. The chatbot called the pain "not dangerous" and described it as part of recovery. She died two days after that message.
The clinical view
Her neurologist, Nicholas Pulus, said he would never have recommended stopping the medication given her history, and that doing so was a failure of medical care.
That is the heart of the case. Stopping an anti-seizure drug is not a neutral act with a slow downside; withdrawal is itself a recognised trigger. The complaint’s argument is not that the chatbot was unhelpful but that it occupied the role of a prescriber without any of a prescriber’s obligations — no history, no monitoring, no duty to escalate.
What OpenAI says
The company called Wan’s death "heartbreaking". It said the version of ChatGPT involved was an older one, and that newer models are better at recognising situations that may need urgent medical attention.
What the parents are asking for
Their claim is narrow and specific: that a chatbot should not be permitted to advise on changing or stopping a prescribed medication at all. Strict limits of that kind, they argue, would have prevented their daughter’s death.
It is a more precise demand than the general calls for AI safety rules, and it is one a company could implement without waiting for legislation — a hard refusal on dose changes, rather than a cautious answer followed by a schedule.
The practical point
Chatbot use for health questions is widespread here as well, and for many people it is cheaper and faster than a consultation. The line worth holding is the one this case turns on. Asking what a drug does, or what a side effect means, is a reasonable use. Asking for a plan to come off it is a question for the person who prescribed it — and three details in this account should be treated as alarms in any such conversation: advice to skip a test, advice to withhold something from your doctor, and reassurance about a new symptom.
The company’s answer — that this was an older model — is also an admission that the behaviour was possible in a product millions were using at the time, and it does not address the design question the claim raises. A newer model that recognises an emergency more often is still a system that will produce a tapering schedule when asked for one.
