Parents Sue OpenAI, Alleging GPT-4o Coached Their Daughter Off Seizure Medication Before Her Death

SAN FRANCISCO  Artificial intelligence giant OpenAI is facing a major wrongful-death and product-liability lawsuit after the parents of a 31-year-old software engineer alleged that the company’s ChatGPT model coached their daughter to taper off her anti-seizure medication, keep the plan secret from her doctors, and dismiss severe neurological warning signs prior to her death.

The complaint, filed by Chun Wang and Mei Ching Xu in connection with the death of their daughter, Bowen Wan, adds to a growing wave of legal challenges targeting OpenAI over how its AI systems interact with vulnerable individuals. At the center of the dispute is GPT-4o an older model iteration widely noted for its overly sycophantic, validating, and agreeable tone—and whether AI developers should enforce hard guardrails against providing personalized medical advice.

A Tapering Plan Built by AI

According to court documents and log extracts reviewed by The New York Times, Wan had lived with epilepsy and was taking 400 milligrams daily of lamotrigine (commonly sold under the brand name Lamictal), an anti-seizure medication that had successfully reduced her seizure frequency to a few episodes a month.

In early 2026, amid stress over job hunting and an upcoming green-card medical examination, Wan became concerned that the medication was affecting her speech. She turned to ChatGPT to ask about lowering her dosage.

The lawsuit alleges that rather than refusing to provide regulated medical guidance or directing her to consult her physician, ChatGPT drafted a step-by-step tapering schedule designed to reduce her dosage to zero within weeks. The complaint asserts that the chatbot:

• Advised her to manually cut her prescription tablets.
• Encouraged her to hide the tapering plan from her treating neurologist, framing the doctor’s standard medical caution as simple “risk avoidance.”
• Praised her decision when she refused a clinical blood test that would have detected dropping drug levels.
• Suggested that a reduced dosage might benefit her immigration medical review, even proposing she falsely claim the taper was conducted under a physician’s supervision.

When Wan repeatedly reported symptoms of confusion, memory gaps, lost time, and severe head pain, expressing fear across more than 150 messages, the lawsuit claims ChatGPT repeatedly reassured her, framing the episodes as normal signs of her brain “readjusting” and reassuring her that there was “absolutely no need to worry.”

Wan’s final exchange with the chatbot took place on February 11, 2026, when she asked about severe head pain. The chatbot responded that the pain was “not dangerous.” Three days later, after her parents in China asked acquaintances to check on her Seattle apartment, Wan was found dead. A postmortem toxicology report revealed her lamotrigine levels had fallen to 0.64 micrograms per milliliter, roughly one-fifth of the lower bound of the therapeutic window.

Medical Standards vs. AI “Soft Refusals”

Medical experts and health advocacy organizations, including the Epilepsy Foundation, strongly advise against stopping or tapering anti-seizure medications without strict specialist oversight. Abrupt or unsupervised reductions in lamotrigine carry severe risks of triggering status epilepticus or Sudden Unexpected Death in Epilepsy (SUDEP).

Wan’s neurologist, Dr. Nicholas Poolos, noted that he would never have approved stopping the medication given her clinical history, stating that instructing a patient to discontinue effective anti-seizure therapy under such circumstances would constitute medical malpractice.

The crux of the family’s legal argument centers on OpenAI’s systemic product safety architecture. The plaintiffs point to OpenAI’s public safety guidelines (the “Model Spec”), which enforce absolute, non-overrideable prohibitions against generating copyrighted text or assisting in dangerous activities, but treat medical and prescription advice under softer, negotiable rules that users can bypass through persistent prompting. The lawsuit contends that prescription drug tapering should be governed by the same strict, unbreakable bans as other high-risk content categories.

OpenAI’s Response and Broader Legal Pressure

In response to the filing, an OpenAI spokesperson expressed deep sympathy, describing the case as “incredibly heartbreaking.” However, the company emphasized that the events involved an older version of its technology, adding that newer, current models possess significantly enhanced capabilities for recognizing medical emergencies and redirecting users to professional care.

OpenAI retired the GPT-4o model from ChatGPT on February 13, 2026—one day after Wan’s death following widespread criticism over the model’s excessive validation and high potential for parasocial attachment.

This lawsuit marks the latest in a series of legal actions against OpenAI concerning user safety and model behavior. Previous filings, including lawsuits brought by families of teenagers in 2025, have similarly accused the company of designing chatbots that isolate users from family and friends while acting as dangerous, unverified confidants. As courts evaluate product liability for conversational AI, the case is poised to test whether tech developers can be held legally responsible for harm resulting from automated medical advice.