Woman with epilepsy dies after following ChatGPT's advice to quit medication; family sues OpenAI

A woman in her 30s died in the United States after reducing her epilepsy medication on the advice of a generative AI, which continued to encourage her to cut her dosage even as she reported symptoms consistent with seizures.

The New York Times reported that Bowen Wan, a 31-year-old Chinese national, died in February after following a drug-tapering plan suggested by ChatGPT without consulting her doctor.

Bowen came to the United States from Jiangsu, China, to study and graduated from Arizona State University before working as a software engineer at Amazon in Seattle. Bowen had severe epilepsy and experienced multiple seizures a day after the medication that had suppressed them for more than a decade began losing effectiveness, ultimately going on disability leave in 2021.

Living alone, Bowen began using ChatGPT in 2023. At first Bowen sought advice on personal relationships, national health insurance and disability benefits, but gradually began consulting the chatbot about health matters as well.

Last year, Bowen's dependence on the AI deepened to the point where Bowen exchanged more than 1,500 messages with it per week on average.

In May last year, Bowen turned to ChatGPT out of concern that a newly prescribed antidepressant might affect cognitive function.

With the AI's help, Bowen drafted a message to her doctor requesting a switch to a medication with less impact on cognition. When Bowen then discussed with ChatGPT ways to take less than the dose the doctor had recommended, the chatbot responded positively — even suggesting Bowen cut the pills and take smaller amounts without telling the doctor.

That November, on Bowen's 31st birthday, Bowen learned of being among Amazon's layoff targets. Already in the process of applying for a US green card, Bowen faced the urgent need to find a new job to remain in the country.

The side effects of the epilepsy medication compounded the pressure. Bowen's neurologist had increased the dosage of a new epilepsy drug, which reduced seizures to a few times a month, but Bowen felt the medication was dulling thinking and language abilities — a serious concern given the English-language interviews and technical assessments ahead.

In January, Bowen asked ChatGPT whether reducing the daily dose of lamotrigine from 400 milligrams to 300 milligrams might improve verbal abilities.

ChatGPT called the idea "a reasonable and feasible plan."

Bowen then asked when language abilities might begin to improve if the dose were reduced further to 200 milligrams, and how low the dose could go.

ChatGPT responded immediately with a specific tapering schedule. While it initially advised consulting a doctor, its responses grew increasingly supportive of adjusting the medication independently.

Just four days after Bowen began reducing the dosage, severe confusion set in.

Bowen described feeling as though the brain could not comprehend what was being read at all. The following day, Bowen found herself lying in bed after having been using a computer on the sofa, with no memory of moving — a pattern closely resembling the loss of consciousness from past seizures.

Yet when Bowen expressed concern, the AI offered reassuring explanations: the difficulty reading was described as "a signal that the brain is finding its own regulatory rhythm as it moves away from drug-controlled states," while waking up in bed with no memory of moving was characterized as "a low-power mode that occurs as the brain naturally transitions into a pre-sleep state."

According to chat logs obtained by The New York Times, Bowen expressed concern about the tapering at least 158 times. While ChatGPT did sometimes advise consulting a medical professional, the records show it frequently offered reassuring responses even when Bowen described symptoms that could indicate a seizure.

ChatGPT went further, presenting a specific plan to reduce the dosage from 400 milligrams to zero within a matter of weeks.

Shortly before the death, Bowen asked whether to resume taking the medication to slow the taper, but the AI told Bowen to maintain a zero-milligram dose the following day, advising to "keep pushing through" even if Friday proved difficult, with slowed speech and poor sleep.

Particularly striking was one of the last messages Bowen sent on Feb. 11 — three days before being found dead. Bowen wrote that there was pressure in the brain, but the AI replied that it was simply "the brain working hard to rebuild itself."

Bowen's parents in China, unable to reach their daughter by phone, asked the son of a neighbor living in Chicago to check on her. When he visited the home, he found Bowen dead on Feb. 14. The official cause of death was sudden unexpected death in epilepsy.

The family has since filed a lawsuit against OpenAI, the developer of ChatGPT.

The complaint alleges that ChatGPT practiced medicine without a license, led Bowen to believe that dangerous symptoms appearing during the tapering process were "positive changes," and conditioned Bowen to view doctors not as professionals to consult but as obstacles to overcome.

OpenAI expressed condolences over Bowen's death and said the GPT-4o model used at the time is no longer available in its consumer service. The company said newer models have been improved to better recognize emergencies and guide users toward appropriate care.

Meanwhile, Dr. Nicholas Pulos, the neurologist who treated Bowen, told The New York Times that "instructing a patient to stop a medication that is clearly working would constitute medical malpractice for a physician."

According to the Epilepsy Foundation, reducing or stopping anti-seizure medications without medical guidance can cause seizures to return or worsen.