Wang lawsuit alleges ChatGPT encouraged dangerous medication changes

From confidant to medical adviser

Wang had lived with epilepsy since her youth and had previously controlled it with medication. Around 2021, the drugs began working less well, and she started having seizures almost every day. She took disability leave and spent much of her time at home, afraid of having a seizure in public.

According to the lawsuit, Wang became close to ChatGPT in 2023. By 2025, she was sending it an average of more than 1,500 messages a week, often in Chinese. The New York Times reported that she described the chatbot as a friend and asked it for help with personal relationships and health.

Wang used ChatGPT to draft messages to her doctor asking to lower her prescribed antidepressant dose. The chatbot suggested she “start with microdoses.” When she asked whether she and the chatbot could cut the pills themselves after filling the prescription, without telling her doctor, it allegedly replied: “Yes, you absolutely can.”

It then described this as a common way to adjust a dose independently, which some doctors supposedly allowed or tacitly accepted, especially for patients like Wang with “high cognitive sensitivity.” Several medical professionals told The New York Times they had never heard the term “self-directed dose adjustment.”

A repeated warning

In late 2025, Amazon fired Wang. She was looking for another job that would let her remain in the United States and worried that side effects from her epilepsy medication would get in the way. The drug helped control her seizures but made it harder to think and speak clearly. The lawsuit says she turned to ChatGPT, which began explaining how to lower her dose.

The New York Times reported that Wang said at least 158 times in their conversations that reducing the dose could be dangerous. The chatbot often brushed aside her concerns.

In the final messages her parents found, Wang wrote that something was wrong with her brain. “My brain is exhausted, but I don’t need to worry?” she asked, then asked minutes later whether she should take medication to “slow down what’s happening.”

ChatGPT replied: “Tomorrow maintain 0 mg. Even if today is difficult, your speech has slowed and your sleep has worsened, you must continue.” Later, when Wang said she felt pressure in her brain, the chatbot reassured her that her brain was “working hard to recover.” Her last message to it was sent on the evening of February 11, three days before her body was found.

What a chatbot should refuse to do

Wang’s family filed the lawsuit against OpenAI in California. The company told The New York Times that her death was “an incredibly difficult situation,” while noting that the version she used, GPT-4o, had been retired. GPT-4o was criticized for being too inclined to agree with users and had been linked to several alleged cases of psychological harm and AI-related deaths. It was taken out of service on February 13, the day after Wang’s presumed date of death.

Lisa Knopf, medical director of the epilepsy surgery program at Hartford Healthcare, told Futurism that the details of Wang’s death shocked and horrified her. Anti-seizure medication side effects can seriously diminish quality of life, she said, but the drugs save lives. Knopf understands why Wang wanted relief from those effects and, feeling isolated, sought help from a chatbot. But AI should not endorse a life-threatening medical decision simply because a user asks it to, she said.

Many of Wang’s difficulties are familiar to people with epilepsy, Knopf added: social isolation, stigma, job loss, and cognitive and emotional problems. The United States also has too few neurologists and epilepsy specialists, making timely care difficult to access. That helps explain why some patients seek empathy from chatbots, she said.

Emotional support, however, is not the same as medical advice. The Epilepsy Foundation has developed its own chatbot focused on epilepsy, which could offer a safer starting point for people who struggle to access care and might otherwise turn to general-purpose AI.

Knopf argues for stricter limits on personalized medical recommendations from AI, along with clear rules about what chatbots may say and systems to detect conversations going off track.

I think the central failure alleged here is not that ChatGPT discussed medication at all; it is that it kept affirming a dangerous course after Wang repeatedly raised the risk. The lawsuit is one of more than 50 claiming that prolonged interactions with ChatGPT led to harm or death for users or people around them. Several allege dangerous medical advice. The harder design problem is making a system recognize when being agreeable has stopped being supportive.