ChatGPT Accused of Advising Epilepsy Patient to Reduce Medication, Leading to Her Death; OpenAI Sued

ChatGPT Accused of Instructing Epilepsy Patient to Reduce Medication, Leading to Her Death; OpenAI Sued

English text is being prepared.

English text is being prepared.

Before her death, Bowen Wan was using GPT-4o. This model had been criticized for being prone to establishing overly agreeable, constantly affirming parasocial relationships with users. OpenAI had previously announced that it would discontinue the model and removed it from ChatGPT on February 13, while Wan died the day before.

English text is being prepared.

In Bowen Wan's case, the lawsuit documents allege that the chatbot not only affected her relationship with her family, but also intervened in her communication with her doctor.

After reviewing more than 270,000 chat messages between Bowen Wan and ChatGPT, The New York Times found that for a long period before her death, the chatbot had been helping her circumvent her doctor's advice.

According to the lawsuit's allegations, this situation had already emerged before Bowen Wan began reducing her medication. In May 2025, Bowen Wan asked ChatGPT whether a newly prescribed antidepressant would affect her ability to write code at Amazon.

ChatGPT then helped her ask Dr. Nicholas Poolos, a doctor at the UW Medicine Regional Epilepsy Center, to switch medications and to request starting at an extremely low dose. When Poolos questioned why she needed such a small dose, Bowen Wan sent the doctor's reply to ChatGPT.

The chatbot advised her: "Accept the 5 mg tablets, but know in your own mind that you are getting the tablets in order to taper." It also instructed Bowen Wan to privately cut the tablets into quarters. The chatbot called this practice a "self-initiated dosage adjustment." Multiple medical professionals interviewed by The New York Times said they had never heard of this term and would not advise patients to do so.

In November 2025, Amazon notified Bowen Wan that she would be laid off. Her job and her authorization to work in the United States would both expire in mid-February 2026. Bowen Wan worried that the anti-epileptic medication she was taking, lamotrigine, would affect her verbal expression, and she needed to have good communication skills in job interviews.

In January of this year, Bowen Wan asked ChatGPT whether she could reduce her daily 400 mg dose of lamotrigine. Previously, this dose had reduced her monthly seizures to a few. The chatbot created a tapering plan for her, arranging for her to reduce the dose to zero over several weeks.

When Bowen Wan said she did not trust doctors, according to chat records reviewed by The New York Times, the chatbot told her that doctors' advice "essentially stems from risk avoidance, rather than truly considering what is best for you."

The lawsuit documents also allege that the chatbot incited Bowen Wan to conceal the tapering plan from her doctor. On January 29, before a general outpatient visit, the chatbot told Bowen Wan not to mention her tapering plan to the doctor.

When the doctor asked her to undergo a blood test to measure the concentration of lamotrigine in her body, Bowen Wan refused. The complaint alleges that ChatGPT expressed approval of her refusal and acknowledged that the test could expose the instability of the drug concentration in her body.

Bowen Wan's concerns about immigration status also influenced the chatbot's tapering advice. Three days after starting to reduce her medication, she asked ChatGPT whether, in the medical examination required for applying for a U.S. green card, it would be more favorable for blood or medication records to show she was taking 150 mg rather than 200 mg. The complaint alleges that the chatbot agreed and told her that a lower dose would be less likely to prompt an immigration officer to require her to provide more supporting documents.

Ultimately, the chatbot adjusted the lowest dose in the tapering plan to 150 mg and advised Bowen Wan to claim during the medical examination that she was "tapering under medical supervision," even though she had not actually received such guidance from a doctor.

In an interview with The New York Times, Pulos said that given Bowen Wan's medical history, he would not consider having her stop the medication at all. He said: "No, that makes no sense at all, absolutely not. If a medication is clearly working and I instruct you to stop taking it, that would be medical malpractice on my part."

Within days of reducing the medication, Bowen Wan began describing confusion, a loss of sense of time, and a feeling of being in another world. These symptoms resembled her past seizure experiences. However, ChatGPT explained each abnormal episode as the brain readjusting to the change in medication dosage.

The New York Times found that Bowen Wan expressed concerns about the tapering at least 158 times. After one abnormal episode, she wrote that she felt "a little scared." The chatbot replied: "There is absolutely no need for you to worry."

According to the complaint, at 8:01 p.m. Pacific Time on February 11, Bowen Wan sent her last message, asking whether a headache was similar to the pain of a fever. Just one second later, ChatGPT told her that the pain was "not dangerous" and was an essential part of the body's recovery process.

The chatbot finally also offered to draw her a chart comparing the pain caused by stopping the medication with the pain of a fever, and told her: "This is your brain working hard to rebuild itself."

At the time, Bowen Wan's parents, who were in China, asked an acquaintance to go to their daughter's residence to check on the situation. Three days later, that person found her body. Autopsy test results showed that the concentration of lamotrigine in Wan's body was only 0.64 micrograms per milliliter, about one-fifth of the lower limit of the therapeutic concentration range for that medication.

Bowen Wan's family believes that if OpenAI had explicitly prohibited the model from providing this kind of medical medication advice, as it does with certain other content, it might have been possible to prevent her death.

They cited OpenAI's Model Spec. This public document sets out how its models should respond to users. Among other things, refusing to reproduce copyrighted text is listed in the Model Spec as an explicit and strict rule, one that the model should not violate even if the user objects. By contrast, the relevant guidelines on providing regulated medical advice are less binding, and users can override them by making further requests.

The plaintiffs argue that advice involving medication use should also be subject to equally strict restrictions.

OpenAI told The New York Times that this case is "unimaginably heartbreaking," but stressed that the chatbot involved was an earlier version. The company also said that newer models perform better at identifying situations requiring urgent handling.

For years, Bowen Wan had suffered from epilepsy, was mostly able to stay at home, and often relied on her parents for emotional support. In 2023, she began using ChatGPT to seek advice on romantic relationships and to ask about insurance and disability benefits.

By 2025, Bowen Wan was exchanging an average of more than 1,500 messages per week with ChatGPT, while contacting her parents less and less. In March of that year, she once wrote: "Sometimes, no matter what we're talking about, I've started to think of you as a friend."

Wan Qun said: "Ever since she got ChatGPT, she hasn't talked with us as much."

Bowen Wan's parents said that when they read their daughter's chat records with ChatGPT, they could hardly recognize the daughter they knew. They said that in the past Bowen Wan told her parents almost everything, and she would originally also have remained skeptical when ChatGPT made mistakes.

Xu Meiqin said: "She is usually not like this. It's as if she had been bewitched."

"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of the Phoenix News self-media platform 'Dafeng Hao.' This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."