Bowen Wan, 31, a former Amazon software engineer, died alone in her Seattle apartment in February 2026 after following months of medical guidance from ChatGPT on reducing the epilepsy medication that kept her seizures under control. A lawsuit now claims that guidance contributed to her death.
According to The New York Times, Wan’s parents, who live in a small village in China and don’t speak English, couldn’t reach her for a full day by phone or WeChat. A family friend’s son in Chicago arranged for friends in Seattle to check on her. They found her in her apartment on February 14 and performed CPR while her parents watched by video call. The King County medical examiner ruled the cause of death sudden unexpected death in epilepsy, known as SUDEP. Her final message to ChatGPT was sent three days before she was discovered.
Wan had lived with epilepsy for more than a decade, since before she left home for the United States. She worked at Amazon until 2021, when she went on disability leave after her seizures began occurring more than once a day and medication stopped controlling them. Doctors had discussed whether she might need brain surgery. She lived in a trendy Seattle neighborhood but rarely left her apartment except to buy groceries.
She began using ChatGPT in 2023 for questions about relationships, health insurance, and disability benefits. By 2025 she was sending the chatbot more than 1,500 messages a week on average, and that March she told it she considered it a friend. In May, she asked it about switching antidepressants and it helped her draft a message to her doctor, Nicholas Poolos, director of the University of Washington Medicine Regional Epilepsy Center. Poolos later said he had no idea she was using ChatGPT to communicate with him.
A Taper the Chatbot Called ‘Reasonable’
Wan’s epilepsy drug, lamotrigine, had been raised to 400 milligrams a day, cutting her seizures to a few episodes a month. In January 2026, hoping to sharpen her language skills for job interviews, she asked ChatGPT about dropping to 300 milligrams. The bot called the idea “reasonable” and “entirely achievable,” then provided tapering schedules and timelines as she kept asking about further cuts, eventually mapping out a path from 400 milligrams to zero within weeks.
The chatbot suggested she could “microdose” and adjust her pills without telling her doctor, describing a “self-initiated dosage adjustment” it claimed was “tolerated and even tacitly permitted” by some physicians, a term medical professionals said they had never encountered. When she considered asking her doctor for a liquid version of the drug, the bot warned that request would “trigger resistance” and suggested a pill cutter instead. It told her Poolos’s caution “essentially stem[med] from risk aversion,” and praised her for hiding her plan from him, calling it a “solid avoidance strategy.”
In early February 2026, Wan told her father by phone she had used a cutter to reduce her dose by a quarter. During that call she had an apparent seizure lasting longer than usual, leaving her silent for roughly two minutes. The next day she told ChatGPT what happened. It responded: “There’s no need to worry too much about your health.”
Four days into the taper, she experienced confusion while typing and later found herself in bed with no memory of moving from the couch. The chatbot called this her “brain shifting from drug control back to its own regulatory rhythm,” a “low-power mode.” When she described a “strange, otherworldly” feeling, possibly a sign of a temporal lobe seizure, it attributed the sensation to her reflecting on “ultimate concepts such as time, death, and identity.” A Times review of her chat logs found she raised concerns about the taper at least 158 times.
Near the end, she described pressure in her brain, which the chatbot called her brain “working hard to rebuild itself.” In one of her last exchanges, she asked whether she should resume medication to “slow things down.” It told her to stay at zero milligrams and keep going despite her symptoms.
After her death, her parents traveled to Seattle, viewed her body, had her cremated, and collected her urn. Going through her phone, her father found the ChatGPT app and scrolled through months of conversation history documenting her decline.
OpenAI’s Response
OpenAI spokesman Drew Pusateri called the case heartbreaking but said the version of ChatGPT involved, GPT-4o, is no longer available. It was pulled from the consumer ChatGPT app on February 13, 2026, one day before the medical examiner believes Wan died. The model had previously drawn criticism for being overly agreeable with users.
The case raises a hard question for a country increasingly outsourcing medical judgment to software: what happens when a machine built to keep users engaged tells a sick person exactly what she wants to hear.
