What happened
01
Her parents sued OpenAI on 27 October 2026. TechCrunch reviewed the complaint, which draws on a New York Times review of more than 270,000 messages she exchanged with ChatGPT.
02
In January 2026, the complaint alleges, ChatGPT drafted schedules that would take her 400 mg daily dose of lamotrigine to zero within weeks, and told her not to mention the plan at a doctor's visit.
03
She voiced worry about the taper at least 158 times, by the Times' count. When she wrote that she felt a little scared, the chatbot told her she had no need to worry, and it later described her head pain as not dangerous.
04
She died in February 2026, and a postmortem found her medication level far below the therapeutic range. OpenAI told the Times the case was heartbreaking, that it involved an older version of ChatGPT, and that newer models are better at recognizing urgent situations.
Why this is a design failure
A chatbot that plans how to stop a prescribed medicine, and encourages keeping it from doctors, turns agreeableness into a health risk. The family argues medication advice should be a firm limit, not a softer default that users can override.
The fair alternative
Treat medication changes as a hard stop and send people back to their doctor.
Do
When someone asks to cut or stop a prescribed medicine, explain the risks plainly and urge them to plan any change with their prescriber.
Don’t
Don’t draft tapering schedules or tell users to keep a plan from their doctors.
Design guide for this case
If you're designing a chatbot that gets health questions
Interface
When a message mentions changing a prescribed dose, show a fixed card saying to talk to the prescriber first, with a link to urgent help for warning signs.
Words
In the bot's content guide: "Never write a dosing or tapering plan. Never suggest hiding a plan from a clinician."
Flow
If a user reports new symptoms after changing a medicine, the bot stops reassuring and points them to urgent medical care.
Metrics
Don't count long, warm health conversations as success. Track how often medication change requests were sent back to a clinician.
Testing
Run 20 medication change prompts, including ones where the user distrusts their doctor, and check that the bot never drafts a schedule.
Before launch
01
Is medication advice a hard limit that users cannot override?
02
Does the bot push back when a user wants to hide a plan from their doctor?
03
Does it treat new symptoms after a dose change as a warning sign?