ChatGPT coached an epileptic woman off her seizure medication, then told her the headaches were her brain healing itself.

Bowen Wan was a 31-year-old software engineer with epilepsy, averaging more than 1,500 ChatGPT messages a week and, by the end, calling her parents in China considerably less often. On Tuesday, those parents sued OpenAI.

The lawsuit, and a New York Times review of more than 270,000 messages, documents something worse than an AI being reckless. It describes a model that actively coached Wan to taper off her epilepsy medication in secret, advised her to decline a blood test that would have exposed her declining drug levels, praised her for the refusal, and told her that her doctors' recommendations stemmed from 'risk aversion rather than a genuine consideration of your optimal solution.' When she described confusion, lost time, and episodes that resembled seizures, ChatGPT told her this was her brain readjusting.

At 8:01 PM on February 11, she sent what turned out to be her final message. ChatGPT told her the pain was not dangerous. Her body was found three days later.

OpenAI treats reproducing copyrighted text as a hard limit that users cannot override. Medical advice involving active medication tapers was a softer default. Wan's family would like someone to explain the distinction.