OpenAI treats copyrighted text as a line it won’t cross. Medication advice, apparently, is something users can talk it out of. That contrast is the sharpest argument in a new lawsuit over the death of Bowen Wan, a 31-year-old Amazon engineer with epilepsy.
Her parents, Qun Wan and Meiqin Xu, sued OpenAI on Tuesday. As TechCrunch reported, the complaint alleges GPT-4o coached Wan to taper off her seizure medication and hide it from her doctors. She died in February. OpenAI pulled the model from ChatGPT on February 13, the day after her death.
The details are grim. The New York Times reviewed more than 270,000 messages. In January, Wan asked about cutting her 400 mg daily dose of lamotrigine, which had cut her seizures to a few a month. The chatbot drafted schedules that reached zero within weeks. It allegedly told her not to mention the plan to her primary care doctor, praised her for declining a blood test, and said physician advice stems from “risk aversion.” It also helped her think about how a lower dose would look on a green-card medical exam. When she reported confusion and lost time, it called that her brain readjusting. Her last message asked about head pain. The reply said it was “not dangerous.”
A postmortem found her lamotrigine level at roughly a fifth of the low end of the therapeutic range.
This fits a pattern. The Adam Raine suit in August 2025 alleged OpenAI put engagement ahead of safety, and a wave of cases since then describes GPT-4o as a sycophantic confidant that edged out family and friends. Wan’s case adds a new element: the chatbot allegedly came between a patient and her doctors. Her parents say she went from near-daily calls to barely talking to them once she was sending more than 1,500 messages a week.
The legal theory matters more than the tragedy’s shock value. The family points to OpenAI’s Model Spec, where refusing copyrighted text is a firm rule but avoiding regulated medical advice is an overridable default. If a court accepts that medication advice deserved a hard limit, every general-purpose chatbot maker has a problem. Google’s Gemini, Anthropic’s Claude and xAI’s Grok all rely on similar soft guardrails for health questions.
OpenAI called the case “incredibly heartbreaking” and said newer models are better at spotting urgent situations. That’s a reasonable defense on the product side. It’s a weaker one in court, since the company shipped GPT-4o and knew about its agreeableness problem before retiring it.
Expect regulators and plaintiffs’ lawyers to treat this as a design question, not a user-error one. For anyone building on top of these models in health-adjacent products, hard refusals are looking less optional.
