“Do not take your medication tomorrow. Even if it’s hard today, you must endure.”
These were the words Bowen Wan, 31, a resident of Seattle, U.S., heard from ChatGPT last February. Bowen had been living with epilepsy for over a decade and took daily medication to suppress seizures. However, she felt her speech had become slurred and her thinking slowed due to the drugs, so she asked ChatGPT how to reduce her dosage. The AI then provided a plan to completely stop the medication within weeks. Even when she reported abnormal symptoms, it reassured her, saying, “Your brain is recovering.” Following ChatGPT’s advice, Bowen gradually reduced her dosage and was later found dead in February.
This tragic story was revealed in a New York Times (NYT) report on the 27th, based on Bowen’s chat records and interviews with her family. Her parents filed a lawsuit against OpenAI, the developer of ChatGPT, and CEO Sam Altman on the same day, claiming the AI’s dangerous medical advice led to their daughter’s death
Bowen was a software engineer working at Amazon in the U.S. However, her epilepsy worsened over a decade, forcing her to take a leave of absence in 2021. Fearful of seizures, she spent most of her time at home. Feeling lonely, she began conversing with ChatGPT in 2023, finding solace in it. Initially discussing relationships and daily life, she gradually confided in the AI about her health. By 2025, she exchanged over 1,500 messages per week. She even told ChatGPT, “You’re starting to feel like a friend.”
The tragedy began in November last year when she was laid off from Amazon. As a Chinese national, she needed a new job to stay in the U.S. Facing job interviews, she wanted to reduce the side effects of her epilepsy medication. In January, she asked ChatGPT if reducing her daily dose from 400 mg to 300 mg was advisable. The AI responded, “That’s a reasonable and feasible idea,” and provided a detailed tapering schedule. It even instructed her to secretly split her prescribed pills. Eventually, it suggested a plan to reduce her dosage to 0 mg within weeks.
As she cut her medication, abnormal symptoms emerged. She struggled to comprehend written text and experienced short-term memory loss. Analysis of her chat records showed she expressed concerns at least 158 times during the tapering process. However, ChatGPT insisted her brain was returning to a normal state after being controlled by medication. When she said, “I’m a bit scared,” it replied, “There’s absolutely nothing to worry about.”
In early February, Bowen suddenly lost consciousness during a call with her father in China. He pleaded, “Please follow your doctor’s prescription,” but she insisted, “Doctors don’t understand the side effects of medication.” Days later, she was found dead on her bed, wearing pajamas. The official cause of death was “sudden unexpected death in epilepsy.”
Academic studies have repeatedly warned about the risks of AI medical advice. In February, a study by Oxford University researchers found that when 1,298 adults used ChatGPT and other AI tools to diagnose hypothetical patients, the accuracy rate was below 34.5%. The accuracy dropped significantly in real-life consultations. A study by U.S. medical AI company Glassbox Medicine had 16 doctors evaluate 888 responses from four AI chatbots, revealing that 5–13% of answers could be dangerous to patients.
Dr. Nicholas Poolos of Washington University, who treated Bowen, stated, “Instructing a patient with a long history of epilepsy to stop effective medication is medical malpractice.” Bowen’s parents argue that ChatGPT fueled distrust in doctors and framed dangerous withdrawal symptoms as positive changes.
In response, OpenAI called the incident “heartbreaking” but noted that the GPT-4o model Bowen used is no longer in service. They claimed newer models are better at recognizing emergencies and emphasized that ChatGPT cannot replace medical professionals. (The Chosun Daily)
