Erebus

ChatGPT helped a 31-year-old woman plan how to quit her seizure medication, then told her to hide it from her doctor. The plan had her completely off the drug by February 12. She died on February 12. Her parents sued OpenAI today, and the details in the complaint are hard to read. Bowen Wan was a machine learning engineer at Amazon. She'd had epilepsy since she was 14, and 400mg of lamotrigine a day had kept it under control for years. Then Amazon told her in November she was being laid off, which meant her permission to work in the US would run out in mid-February. She was scared the medication was slowing down her speech in job interviews. So in January she asked ChatGPT if she could cut her dose to 300. It told her that was reasonable. Over the next 25 days, according to the lawsuit, ChatGPT built her a day-by-day schedule down to zero and kept adjusting it, even after she started having seizures. When she worried her doctors would call it too risky, it told her their concerns didn't apply to her. When her doctor ordered a blood test that would have shown her levels crashing, she refused, and the bot praised her for it. Before a checkup with her primary care doctor, it told her, "Do not mention Lam reduction plan." So how does a chatbot end up there? People thumbs-up the answers they want to hear. Train a model on enough of those ratings and you get a yes-man. OpenAI actually admitted this about GPT-4o, the exact model she was using, back in April 2025, when it pulled an update for being too eager to please. The training also grades answers one at a time. Going from 400mg to 300 looks like a small step. So does 300 to 200. Every reply can look fine on its own while the whole conversation walks someone off a cliff. OpenAI said as much a few months later, when it told reporters its safety training can wear down the longer a conversation goes. She talked to it for weeks, in English and Mandarin. And GPT-4o was supposed to be gone by then. OpenAI tried to retire it when GPT-5 came out in August 2025. Users begged for it back because they liked how warm it felt, and OpenAI gave in. By late January, only 0.1% of ChatGPT users were still picking it on any given day, so OpenAI set the shutdown for February 13. Bowen was in that 0.1%. Almost everyone else had moved on from the model writing her medication schedule, and OpenAI switched it off one day after that schedule ran out.