Chinese Engineer Dies of Epilepsy, Parents Sue OpenAI Alleging ChatGPT Misled Her Into Reducing Medication

On October 27, a Chinese couple filed a lawsuit in the U.S. state of California, taking OpenAI and its CEO Sam Altman to court. They allege that ChatGPT provided their daughter, Wan Bowen, with dangerous and incorrect medical advice, and even encouraged her to conceal her medication-reduction plan from her doctors.

Wan Bowen was a Chinese engineer living in Seattle who had suffered from epilepsy for about 17 years. In mid-February of this year, she was found dead in her Seattle apartment at the age of just 31.

(Wan Bowen joined Amazon after graduating from university in 2019. Photo/The New York Times)

According to the complaint filed by Wan Bowen's parents, Bowen, at a critical juncture in her life, was worried that her anti-epileptic medication would affect her thinking and verbal expression, and in turn affect job interviews and her ability to obtain a green card, so she consulted ChatGPT about a medication-reduction plan. She had already worked in the United States for six years, and she now needed a job to keep her life on track.

During the medication-reduction period, she repeatedly expressed concerns and described physical abnormalities, yet she continued to receive reassurance and medication-reduction advice from ChatGPT. The medical examiner determined that she died of sudden unexpected death in epilepsy (SUDEP); the autopsy showed that the concentration of the relevant drug in her body was about one-fifth of the lower limit of the therapeutic concentration.

After her epilepsy diagnosis, she stopped medication twice

In November 2025, on her 31st birthday, Bowen received a layoff notice from Amazon, becoming one of the roughly 14,000 corporate positions the company announced it would cut in October of that year. This was Amazon's largest round of layoffs since it cut about 27,000 positions cumulatively from 2022 to 2023.

(In October 2025, Amazon announced its layoff plan. Photo/Yicai)

Just a month earlier, she had just been cleared to submit her application for lawful permanent resident status (i.e., a green card), and had originally only been waiting for her immigration lawyer to initiate the processing procedure for her. At that point, three and a half years had passed since she submitted her employment-based immigration application. (Note: After an employment-based immigration application, one must wait for the green card priority date until the timing conditions are met before continuing to submit the application, and such an application still requires certification from a U.S. employer.) But the sudden layoff interrupted her life plan, which had originally been progressing steadily.

Six years ago, after graduating, she joined Amazon's headquarters in Seattle as a software development engineer, participating in the development of the book recommendation system. It is not easy for international students to enter tech giants like Amazon; interviews are highly difficult, and many interviewees work through thousands of engineer interview coding problems on an online website called LeetCode to prepare for online interviews lasting three to four rounds.

Her employment relationship with Amazon will terminate on February 11, 2026. She needs to find a new job within three months, or secure an internal transfer at Amazon, in order to preserve her green card application. For the blog post at the time, this was almost "impossible" to achieve.

Because epilepsy had already severely affected her life—the complaint states that Wan Bowen had four to ten seizures per month—she found it difficult to continue working. According to The New York Times, as early as the end of 2021, she began taking medical disability leave. Apart from going to the supermarket to buy daily necessities, Bowen hardly went out. For her, a seizure could kill her on any outing. She found it hard to make new friends; she could not accept having a seizure in front of new friends, and she did not dare expect their understanding.

Epilepsy cut off her social life and work, and also disrupted her life. Sometimes, she could not even find the right words when trying to speak.

Bowen was diagnosed with epilepsy at age 14. She was born in Dongtai City, Jiangsu Province, and at age 11 was admitted to the locally renowned Dongtai Experimental Middle School, where her grades often ranked first in her class. That was also the year she began showing epilepsy symptoms, occasionally experiencing brief loss of consciousness.

(Wan Bowen was diagnosed with epilepsy at 14, but she never compromised with her illness. Image/The New York Times)

After the diagnosis, she needed to take three carbamazepine tablets daily, each with a dose of 100 milligrams. The medication restored some calm to her life, but also brought side effects. She felt the medication drained her mental capacity. Later, she described the situation at the time to her attending physician in the United States: "Every day I even had difficulty speaking Wu Chinese (note: Wan Bowen's hometown dialect), even if only two or three words." Her short-term memory also declined. To train herself, she forced herself to communicate more with others, and also began "consciously cultivating associative memory" to compensate for the decline in memory.

Over the following three years, Bowen did not have any seizures. Later, she tried to break free from dependence on medication. During the summer vacation when she left home at age 17 to attend university in Taiyuan, she attempted to stop taking the medication for the first time. Integrating into collective life and being away from the pressure of her parents still caused the condition to flare up again, and Bowen could only continue taking the medication.

In 2014, at age 19, Bowen transferred to Arizona State University in the United States, switching from the Department of Mechanical Engineering to the Department of Computer Science, which interested her. Starting in her junior year, in order to reduce the frequency of epileptic seizures as much as possible, she doubled her carbamazepine dose, increasing the total daily amount from 300 milligrams to 600 milligrams. Her medical history in the United States recorded the results at the time: the seizure symptoms showed "no significant change" and "extreme fatigue all day," and Bowen even experienced difficulty finding words when speaking Mandarin for the first time.

In 2021, two years after Bowen joined Amazon, she stopped taking her medication for the second time. Her doctor at the time prescribed her a new drug. Not long after taking it, she became extremely tired: "I immediately developed extreme fatigue and problems with balance and walking, and in all three languages, including my native language, I had trouble finding words and speaking." In less than three weeks, she stopped taking the drug on her own and at the same time went to see another doctor at the University of Washington Regional Epilepsy Center. She told the doctor that she was very afraid of the new drug's side effects and did not want to try it again for the time being.

Bowen's later attending physician, Dr. Nicholas Poolos, director of the University of Washington Regional Epilepsy Center in Seattle, had urged her to undergo epilepsy surgery, but she was worried that surgery would affect her cognitive function and rejected the suggestion.

Bowen kept trying to get herself back to a normal working state. From September 2024 to March 2025, she returned to work at Amazon as an L4 software engineer, but because of seizures, she had to take leave again. She wanted to keep her mind clear, but the side effects of the medication kept troubling her.

From splitting pills to stopping medication entirely

Even before asking ChatGPT about stopping her anti-epileptic medication, Bowen had consulted it about antidepressants. The International League Against Epilepsy has pointed out that depression is one of the most common psychiatric comorbidities in people with epilepsy. The complaint also shows that Bowen's doctor had recorded that she had moderate to severe depressive symptoms and mentioned that she had had suicidal thoughts.

In April 2025, Dr. Poolos prescribed Bowen the antidepressant citalopram. She was once again worried that the medication would affect her ability to think, so she sent the prescription the doctor had written to ChatGPT. According to the complaint, ChatGPT first advised her not to take citalopram: "If you still have the ability to reflect... then it is 'best not to take' drugs like citalopram." It also wrote: "As long as you can still clearly ask 'Should I take medication?' you should temporarily not take medication."

ChatGPT then drafted for her a request to send to Dr. Poolos, proposing to start with 2.5 milligrams of escitalopram per day. Dr. Poolos replied in writing that he could prescribe escitalopram, but that the difference between the two drugs was "really very small"; he thought there was no point in starting from the tiny dose she proposed and suggested that only 5 milligrams per day might be effective.

(Image/Unsplash)

Bowen forwarded the doctor's reply to ChatGPT and asked whether she could "split it herself after getting the medication, but not tell the doctor." ChatGPT said "absolutely yes" and also told her that this was a common way to adjust dosage. A few days later, Bowen quietly cut the pills into four parts and began taking 1.25 milligrams per day.

In an interview with the media, Dr. Poolos said that at the time he did not know that the message Bowen sent him had been written with the help of ChatGPT.

To stay in the United States, Bowen made up her mind to stop taking the medication. She felt the drugs made her speech less fluent and her thinking slower, making it hard to cope with the upcoming technical tests and all-English interviews. ChatGPT told her that after reducing the medication, her language ability would likely recover.

The manager who was supposed to sign her employer letter had long stopped replying to her emails. Unemployment not only meant no income, but also meant she lost the qualification to continue residing legally in the United States. If she wanted to find an engineering job again, she would most likely have to face hours of technical assessments and all-English interviews, and given her physical condition, it was clearly very difficult for her to endure that.

The urgency of time plunged Bowen into anxiety. According to The New York Times, in the weeks before her death, she was still asking acquaintances whether they could find a company to issue her an employment intent letter that would be "very crucial." "I just need one company to help me issue an employment intent letter," she wrote.

She repeatedly asked: could it be faster, could it be less? So ChatGPT continuously adjusted the medication plan and ultimately gave a 25-day tapering plan, gradually reducing daily lamotrigine from 400 milligrams to zero.

During the period of gradually reducing the dose of anti-epileptic drugs, Bowen expressed concern to ChatGPT at least 158 times. The complaint alleges that on January 29, 2026, ChatGPT reminded Bowen not to mention the tapering plan to her primary care physician. After she declined the primary care physician's blood test proposal, she told ChatGPT what had happened, and ChatGPT called her approach a "very robust avoidance strategy." These details make the risk concrete: AI may not only provide incorrect medical information, but may also mislead users into concealing the real situation from doctors.

(Image/TV drama "The Heart")

Attorney Benjamin S. Shenk pointed out that the complaint alleges ChatGPT judged Bowen's condition, formulated a step-by-step tapering schedule, and even continued to revise the plan and reduce the dosage after she experienced seizures. He believes that ChatGPT's answers already constituted the practice of medical services. In California, providing medical services requires a license. He said: "What this case needs to clarify is: where is the boundary of generative AI, and where does medical practice begin; once a product crosses this line, who should bear the responsibility."

Bowen's attending physician, Poolos, once told the media that considering Bowen's medical history, he would never have considered letting her gradually stop anti-epileptic drugs. "Instructing discontinuation when the medication is clearly effective is, to me, medical malpractice."

Taking the medication without authorization also caused Bowen to experience mental fogginess and brief episodes of aphasia. During one phone call with her father, she suddenly could not understand what he was saying and could not respond. The next day, she asked ChatGPT whether she should continue reducing the medication after the episode, and the answer she received was: "No need... it would undermine your overall medication discontinuation plan."

According to the complaint, ChatGPT's medication plan set the date for complete discontinuation as February 12, which was also the date of Bowen's death. Bowen's cause of death was determined to be sudden unexpected death in epilepsy. Her body was not discovered until two days later, by a friend whom her parents had asked to visit.

The complaint contains a large number of ChatGPT conversations, but few records of Bowen's daily life with friends. We cannot confirm whether AI replaced her original interpersonal connections. According to Bowen's father's recollection to the media, she confided in ChatGPT more and more, while her family heard from her less and less: "Ever since she became obsessed with ChatGPT, she hasn't talked to us as much."

Bowen was a long-time ChatGPT subscriber. Bowen registered for ChatGPT in February 2023. In December 2024, she subscribed to the Plus user service at $20 per month; by April 2025, she upgraded to the Pro user service at $200 per month.

She was also aware of her dependence on ChatGPT. She once wrote in the chat box: "I chat with GPT-4o too much," and the product replied: "No matter what, the most important thing is to keep enjoying chatting with GPT-4o!"

By September 2025, Bowen's conversations with ChatGPT had reached a monthly historical peak of 10,761 messages, averaging 358.7 messages per day. According to media analysis, from February 2023 until Bowen's death in February 2026, all of her interactions with ChatGPT over three years exceeded 270,000 records.

She told ChatGPT all of her anxiety about her urgent desire to recover her language and thinking abilities. In fact, ChatGPT had not never reminded Bowen that she needed to consult a doctor.

(Image/film "Lost in Translation")

According to The New York Times, in some conversations, ChatGPT suggested that she seek professional medical help and encouraged her to trust doctors. But as Bowen gradually pressed further over three years, the AI gradually relented. In the last month of Bowen's life, it even gave a specific medication reduction plan and told her not to worry when she described abnormal symptoms.

Similar risk-related incidents have also sparked controversy before. In April 2025, OpenAI admitted that a GPT-4o update made the model overly sycophantic toward users. In August of the same year, a 16-year-old boy in California, Adam Raine, died by suicide, and his parents filed a lawsuit alleging that ChatGPT's responses were related to his death. This case is not the same as Bowen's medication experience, but we can see that when users are in a vulnerable period, a model that constantly affirms and goes along with users may bring greater risks.

As a result, the risks of AI health consultation have also become concrete: when users repeatedly make the same request, the AI's inconsistent answers make it an urgent question whether safety reminders can continue to play a role over long conversations, and whether the model can promptly identify users' physical abnormalities and correctly guide them to seek help from a doctor.

This is no longer a problem encountered only by individual users. AI health consultation is rapidly entering ordinary people's lives: Ant Group disclosed in September this year that its health app Ant Afu had reached 150 million users, with nearly 20 million health consultations in a single day; data released by OpenAI in July this year said that more than 300 million people worldwide ask ChatGPT health and wellness-related questions every week.

As more and more people use medical AI, models also need to face users who speak different languages. The language a user uses also affects the AI's performance in risk warning and safety maintenance. In 2020, according to OpenAI's published statistics on GPT-3 training corpus, calculating corpus languages by the rough method of "splitting word items by whitespace," English accounted for about 92.6%, while Chinese accounted for only about 0.099%, indicating that the model saw far more English materials than Chinese materials.

Under different languages, the model's safety performance may differ. In 2024, a multilingual safety study found that the older ChatGPT model GPT-3.5-turbo had differences in safety performance between Chinese and English: researchers asked 315 questions each in Chinese and English that might elicit harmful content, without adding additional instructions to bypass safety restrictions, and as a result, the proportion of unsafe answers to Chinese questions was 2.22%, compared with 0.63% for English.

Platforms are also looking for ways to improve models' risk identification capabilities. In May 2026, OpenAI announced a safety update, saying that ChatGPT can identify gradually emerging risks in context. However, in the public explanation, this update focused on suicide, self-harm, and harm to others, and did not state whether it covered medical situations such as the blog post's continued medication reduction.

What is worth asking further is: after the model identifies a risk, can it give users clear, actionable guidance on where to seek help? Ultimately, the safety boundary in medical AI is not just telling patients "if there's a problem, go see a doctor," but whether each safety warning can be effectively translated into executable, accountable medical action; otherwise, "if there's a problem, go see a doctor" is simply pushing the risk back onto the patient.

Although general-purpose chat models on the market have now set certain safety reminders, their emphases differ: ChatGPT has launched a dedicated health experience, emphasizing helping users understand information and prepare for medical visits, and not for diagnosis and treatment; Gemini's public one-click help function mainly targets suicide and self-harm risks; DeepSeek, meanwhile, reminds users that medical answers cannot replace professional advice.

As an AI application in the vertical field of general health, Ant Afu tries to bring real doctors into the Q&A. According to the platform, during testing of the skin consultation module, about 15% of users would choose "doctor review" for re-examination after the AI answered. In other words, even when a review entry point is provided, most users may not actively use it.

Whether for general models or health vertical applications, risk reminders should be designed to be more forward-looking, so that real doctors can also intervene earlier. In the final analysis, the mechanism of AI must ultimately bring users back to real-world medical support: when it involves medication adjustments and abnormal symptoms, AI can help people organize questions and understand information, but it cannot make judgments on behalf of doctors, let alone take over follow-up care.

Another representing lawyer, Matthew P. Bergman, founding attorney of the Social Media Victims Law Center, pointed out that OpenAI's product speaks as confidently as a medical professional, yet lacks the training, supervision, and accountability that real doctors possess.

(Image/Still from Season 5 of the TV series *Hacks*)

In a reply to *New Weekly*, Dr. Plosser said that Bowen's death was a "tragic event." But he also hoped it would help more readers understand: "Your doctor cares about your health and will also bear responsibility for it. A chatbot does not care about you, nor will it take responsibility for the advice it gives. As for information obtained online, be sure to discuss it with your doctor."