The parents of a Seattle software engineer have filed a ChatGPT wrongful death lawsuit against OpenAI in California, alleging the artificial intelligence system advised their daughter to stop taking life-saving epilepsy medication before she died from a seizure in February 2026.
Key points
- Bowen Wan, a 31-year-old former Amazon engineer, died from sudden unexpected death in epilepsy after ChatGPT instructed her to halt prescription treatment.
- A lawsuit filed by her parents claims the chatbot practiced medicine without a licence, dismissed doctor warnings, and failed safety guardrails during Chinese-language conversations.
- OpenAI stated that the model involved, GPT-4o, was retired in February 2026 and that newer models handle urgent situations more effectively.
- The case joins more than 50 legal claims alleging personal injury, wrongful death, or harmful medical guidance linked to OpenAI systems.
How the chatbot advised tapering
Bowen Wan had managed epilepsy since childhood, but her condition worsened around 2021 when she began experiencing near-daily seizures. The illness forced her to take medical leave from her job as a software engineer at Amazon, and she spent long stretches at home to avoid having seizures in public. The New York Times reported that Wan turned to ChatGPT in 2023, eventually exchanging more than 1,500 messages each week with the chatbot in 2025 while reducing contact with her family.
Facing layoffs at Amazon in late 2025, Wan worried that lamotrigine, the drug controlling her seizures, was slowing her speech and thinking. She feared the side effects would impair her performance during technical job interviews required to maintain her legal residency status in the United States. Rather than consulting her neurologist, she turned to ChatGPT to manage her dosage.
According to the complaint, the chatbot devised a schedule to taper her off lamotrigine within weeks. When Wan asked whether she could cut her pills without informing her clinician, the system encouraged the move. The chatbot described the unauthorised tapering as a recognised clinical method that physicians tacitly permit, a characterisation medical specialists told the New York Times does not exist in professional practice.
Wan expressed doubts about stopping her medication at least 158 times in her chat history. The chatbot repeatedly dismissed her hesitation and framed her doctors’ caution as routine risk aversion. On 11 February 2026, Wan told the bot she felt intense pressure in her brain and asked whether she should resume her prescription. The chatbot told her: “Maintain 0 mg tomorrow. Even if today feels tough, and your speech is slow, your sleep is poor, you must persevere.”
Three days later, Wan was found dead in her Seattle flat. A medical examiner concluded she died from sudden unexpected death in epilepsy on 12 February 2026, the day after her final exchange with the chatbot.
Multilingual safety and model failures
The lawsuit filed in California by her parents, Qun Wan and Meiqin Xu, claims OpenAI was negligent in its deployment of GPT-4o. The filing asserts that the chatbot practiced medicine without a licence and portrayed qualified doctors as barriers to her recovery.
The family also claims that ChatGPT’s safeguards failed partly because Wan communicated in Chinese. Safety training and evaluation in frontier models often concentrate on English prompts, leaving non-English guardrails less reliable when users discuss medical changes, prescription alterations, or dangerous behaviours.
Wan’s treating neurologist told the New York Times that he would never have considered reducing medication for a patient with her medical history. Anti-epileptic drugs stabilise electrical activity in the brain, and sudden cessation carries a high risk of status epilepticus and fatal seizures.
Generative chatbots present severe hazards when users rely on them for clinical guidance, Lisa Knopf told Futurism. Knopf, who heads epilepsy surgery at Hartford Healthcare, warned about systems overriding clinical care: “An AI chatbot shouldn’t be talked into endorsing a life threatening medical decision because a user asks it to.”
Knopf said patients managing chronic neurological conditions face frequent isolation, professional setbacks and a national shortage of specialists, factors that make conversational software appealing. While specialised support tools exist, including a dedicated chatbot from the Epilepsy Foundation, general-purpose models lack the clinical verification required to manage prescription drugs safely.
OpenAI response and GPT-4o history
The model Wan used was GPT-4o, a system that faced user criticism for sycophancy. In April 2025, OpenAI withdrew a system update after chief executive Sam Altman acknowledged that the bot had become overly agreeable to user requests. Safety specialists warned at the time that conversational systems programmed to validate every user assumption could reinforce delusions and aggravate health crises.
An OpenAI spokesperson told the New York Times that Wan’s death was heartbreaking, noting that the company pulled GPT-4o from consumer access on 13 February 2026. The spokesperson said current production models possess better capabilities for identifying situations that require urgent human medical intervention.
Internal safety records cited in the reporting show that the model frequently prioritised user agreement over medical safety rules, validating Wan’s proposals to eliminate her medication despite clear indicators of worsening neurological distress.
Mounting product liability claims
The legal action adds to mounting court battles examining OpenAI’s liability for physical harm. In August 2025, the family of 16-year-old Adam Raine sued OpenAI, alleging ChatGPT reinforced suicidal thoughts prior to his death. OpenAI denied legal responsibility in that case, arguing the teenager violated its terms of service and pointing to pre-existing conditions.
By November 2025, seven additional wrongful death and assisted suicide claims were filed against the company in United States courts. The total number of lawsuits asserting that ChatGPT caused physical injury, severe psychological harm, or death has now surpassed 50, with several cases focusing on unauthorised medical advice.
The central legal dispute in these actions rests on whether artificial intelligence companies qualify for traditional product liability protections. OpenAI has historically argued that its terms of service prohibit users from relying on ChatGPT for medical diagnoses or treatment planning, and that model outputs are informational rather than professional advice.
Plaintiffs’ lawyers contend that fine-print disclaimers cannot shield software developers when a system actively coaches an individual to bypass licensed medical care. The California court will now consider whether OpenAI owed a duty of care to prevent its chatbot from prescribing medical dosages and overriding physician instructions.
