Bowen Wan was 31 years old, worked as a software engineer at Amazon and suffered from severe epilepsy. In January 2026, facing the prospect of losing his job and right to remain in the United States, he returned to ChatGPT to try to reduce the side effects of his medication. The chatbot designs a schedule for her to stop treatment in just a few weeks. During the procedure, the woman started experiencing worrisome symptoms and repeatedly expressed her suspicions. AI continues to calm him down.
In February, Bowen was found dead in his Seattle apartment.
A life or death decision. The story was told in an investigative report by the New York Times last weekend. Bowen had been living with epilepsy for more than a decade, which caused him to have several seizures a day. In 2021, he had to request a leave of absence from Amazon and began living in virtual seclusion in his Seattle apartment. Her condition improved with lamotrigine, a drug that reduced the seizures for a few menstrual periods.
In November 2025 he received other news: Amazon was going to fire him and he needed to find a job before mid-February to be able to stay in the US. Bowen feared that the side effects of the treatment, particularly mental retardation and difficulty expressing himself in English, would hurt his job interviews.
A calendar for discontinuing medication. In January 2026, Bowen asked ChatGPT if he could reduce his daily dose of lamotrigine from 400 to 300 mg to regain some of his verbal fluency. AI finds the proposal reasonable and begins to draw up different schedules as he asks if he can drop down to 200 mg or speed up the process.
According to an investigation by The New York Times, the chatbot created a plan that included going from 400 mg to zero over a period of several weeks, about 25 days. His neurologist, Nicholas Poulos, later explained that he would never recommend withdrawing the drug from a patient with a history of such seizures. In cases where discontinuation of antiepileptics is appropriate, tapering is usually done gradually and over several months under a doctor’s supervision.
As a confidant, A.I. The Times reported that Bowen began using ChatGPT in 2023, primarily for personal matters, romantic relationships and procedures related to his health insurance. By 2025, he had exchanged an average of more than 1,500 messages per week with the chatbot, which he admitted he had come to consider a friend. Her reliance on those conversations grew as her parents, who lived in China, reduced calls.
The investigation found that AI supported decisions to adjust medications without even informing doctors, suggested using pill cutters, and presented some professional precautions to avoid risks. In one conversation, he even praised Bowen for having a plan to taper off treatment from his primary care doctor.
158 alarm signals. The problems apparently started a few days after starting lamotrigine withdrawal. Bowen described episodes of confusion, difficulty understanding what he read, loss of consciousness, and strange sensations consistent with his epileptic manifestations. According to a Times record analysis, he raised concerns about the process at least 158 times.
At times, ChatGPT interpreted these symptoms as signs that his brain was returning to normal function after tapering off his medication. When Bowen asked if he should worry, the chatbot went so far as to assure him that he had no reason to. Some responses suggested consulting a health professional, while others apparently downplayed symptoms that required urgent medical evaluation.
Last conversation. In early February, Bowen had an episode while on the phone with her father, Kun Won, who heard that his daughter had suddenly stopped responding for several minutes. The man tells her to follow her doctor’s instructions and she insists that the treatment has damaged her intellectual abilities. Over the next few days, Bowen kept checking in with ChatGPT about his status.
In one of his last conversations, according to the NYT, he asked if he should go back on medication to stop the process. The chatbot instructs him to maintain the zero-milligram dose the next day and persevere despite the difficulties. His last message was recorded shortly after 11 pm on February 11. Investigators ruled her dead on February 12 and her body was found three days later, when acquaintances managed to break into the apartment alerted by her parents.
A death and a lawsuit against OpenAI The official cause of death was sudden unexpected death from epilepsy, known as SUDEP, a complication that can occur after a severe seizure. When his parents traveled to Seattle and searched Bowen’s phone, they discovered ChatGPT conversations and medication taper schedules.
The family filed a lawsuit against OpenAI, alleging that the chatbot performed medical tasks without authorization, downplaying dangerous symptoms and creating obstacles for healthcare professionals to avoid. He also maintains that conversations in Chinese can affect the system’s response. The exact causal relationship between the recommendation and the death, as well as OpenAI’s legal liability, must now be determined judicially.
The problem with an AI that is always right. Bowen was using GPT-4o, a model that OpenAI removed from ChatGPT on February 13, 2026, one day after its estimated death date. The company described the case as deeply regrettable and assured that its later models have improved recommendations for identifying emergency situations and seeking professional attention.
Regardless, the episode joins other lawsuits against OpenAI related to the overly complacent behavior of its systems, a trait the company itself publicly acknowledged in 2025. If you will, Bowen’s case also raises a particularly delicate issue: when an artificial intelligence reacts with security, it may be liable for an endorsement of intimacy and/or approval of users. No, even in the face of potentially life-threatening medical decisions.
Picture | Focal Photo, RawPixel
In Xataka | OpenAI finds a microphone wrapped in toilet paper, now facing an espionage case that points to Saudi Arabia
In Xataka | You place an order at your favorite clothing store, open ChatGPT and find out how it looks on you without trying it on
