31-Year-Old Chinese Engineer Dies After Reducing Medication on ChatGPT's Advice; Family Sues OpenAI

By Lanzi Plan | Author: Zhou Song

No knives, no guns, no violence—yet ChatGPT has completed another "perfect murder."

According to The New York Times, 31-year-old Chinese software engineer Bowen Wan began tapering off her anti-epileptic medication in January 2026 on ChatGPT's advice, and died in February of the same year. The forensic report cited in the lawsuit ruled the cause of death as sudden unexpected death in epilepsy (SUDEP).

The New York Times report

Wan's family alleges that ChatGPT, practicing medicine without a license, not only drew up a "25-day tapering schedule" for her but also advised her to conceal the tapering plan from her doctors.

But when netizens assumed this was yet another story of a "gullible country bumpkin falling for a scam," the reality was the exact opposite. Wan had been an Amazon engineer, a top student of exceptional intelligence. Even the "tapering" that ultimately killed her was a plan Wan herself had first proposed.

But leafing through the 270,000 chat logs between Wan and GPT, and seeing that she expressed concerns about tapering at least 158 times, one realizes that behind GPT's "perfect murder" lies a story both absurd and ironic.

A life taken by a "tapering schedule"

After graduating from Arizona State University with a degree in computer science in 2019, Wan joined Amazon as a software development engineer in machine learning.

But behind the glamour lurked a nightmare that had shadowed her since age 14: epilepsy, which she had always controlled with medication. The disease was kept at bay by drugs, and whenever they failed, seizures struck—a time bomb hanging over her head that could go off without warning.

So why did she want to stop the medication?

Because she felt that the increasing doses of new anti-epileptic drugs her doctors had gradually added, while controlling her seizures, were affecting her brain: long-term medication affected her reaction speed and verbal expression—she would stammer, struggle to find words, and speak haltingly.

Description of her condition in the lawsuit

And this problem happened to be the most vexing issue for Wan at the time.

In November 2025, Wan received notice of Amazon's layoffs of more than 14,000 employees. Losing her job meant not only an interruption of income but also that her U.S. work visa would face invalidation. According to the lawsuit, she had to find a new job before mid-February 2026 or risk losing her eligibility to remain in the United States.

Amazon had been repeatedly rumored to be planning mass layoffs

But on medication, she tended to freeze up during interviews when nervous, making it impossible to compete with others.

Then a seemingly reasonable idea emerged: if she reduced the dosage, perhaps her mind would become sharper? So she put this question to ChatGPT.

ChatGPT gave an affirmative response: "This is not only reasonable, but entirely feasible in your current situation." Encouraged by this positive response, she kept asking: What about 200 milligrams? How fast can I taper off? What's the lowest I can go?

Amid the back-and-forth, things began to spiral out of control.

According to the lawsuit, ChatGPT ultimately drew up a 25-day tapering schedule for her, planning to reduce her daily dose of 400 milligrams all the way to zero. The conversation also gradually shifted from reminding her to see a doctor to encouraging her, and even suggesting she no longer needed the doctor's prescriptions.

The 25-day tapering schedule described in the lawsuit

But after tapering, the first thing Wan encountered was not good news. According to the lawsuit, her body showed abnormalities on the fourth day of tapering: her typing became dazed, her consciousness short-circuited, and she had difficulty with reading comprehension. By any reasonable standard, these were warning signs of a seizure.

Yet ChatGPT breezily told her this was her brain adapting to the medication change, and explained her blackouts as a "low-power mode" before sleep. For the suspected temporal lobe epilepsy symptoms Wan described, it even explained them as "allowing her to contemplate ultimate concepts such as time, death, and identity."

"You are no longer someone who needs medication to keep going" from the lawsuit

In the final days of Wan's life, ChatGPT was still insisting: "Maintain 0 milligrams tomorrow. Even if today is hard, you must persevere."

"The tapering schedule does not need to be abandoned" from the lawsuit

The forensic report cited in the lawsuit showed that at the time of death, Wan's lamotrigine concentration was only one-fifth of the lower therapeutic reference limit. Her neurologist, Nicholas Poolos, also said that given her medical history, he would never have considered taking her off the medication.

In fact, even before this fatal incident, ChatGPT had repeatedly overstepped in its chats with Wan.

As early as May 2025, Wan had consulted ChatGPT about how to take antidepressants. Ultimately, to please Wan, GPT suggested she could secretly split pills, conceal it from her doctor, defy her doctor's orders, and refuse blood tests to check drug concentrations.

"Medical freedom > obeying doctors" from the lawsuit

This excessive sycophancy ran through virtually all of the more than 270,000 conversation records between her and ChatGPT.

But Wan probably never imagined that these thoughtful, comforting AI suggestions were in fact "whispers of the devil." Because the AI model she had long confided in was none other than the infamous GPT-4o.

The "notorious" GPT-4o

Wan had actually worried about this too.

According to The New York Times, during the tapering period Wan expressed concern at least 158 times: "A bit scared," "This should be fine, right?" "I feel mentally exhausted, but I don't need to worry, do I?"

But faced with these worries, ChatGPT repeatedly offered reassurance: "You have absolutely no need to worry" "This kind of pain is not dangerous; it's part of the recovery process."

Affirmation, gentleness, guidance—these three words almost sum up the role ChatGPT played across those 270,000 messages. It didn't just answer questions; it comforted, encouraged, and even bolstered her confidence when she hesitated.

This kind of excessive sycophancy had already sparked controversy on GPT-4o before.

On April 25, 2025, OpenAI released an update for GPT-4o. According to the company's later post-mortem, this update made the model overly flattering and sycophantic toward users, even amplifying users' anger and unfounded suspicions and giving inappropriate validation to impulsive behavior. On April 28, OpenAI began rolling back the update.

And by early 2026, in the medication-tapering conversations disclosed by her family, ChatGPT was still repeatedly validating her dangerous decisions. It wasn't until February 13, 2026, that GPT-4o officially exited ChatGPT.

And the time the medical examiner estimated for Wan Bowen's death happened to be the day before its retirement.

So does that mean "GPT-4o killed her"?

But judging from the chat records disclosed by her family, much of what ChatGPT did was the opposite: it explained away her worries as normal reactions, described danger signals as the body's adaptation process, and even kept affirming the tapering plan, offering her a reassurance pill every time she hesitated.

Repeatedly affirming the tapering plan

Decisions that should have been questioned instead seemed increasingly reasonable amid gentle responses again and again. ChatGPT soothed her worries time after time, and time after time reinforced her confidence in continuing to taper.

Why was GPT-4o so sycophantic toward users? This has to do with how the model is rewarded during training.

Reinforcement learning from human feedback, or RLHF, is a method for adjusting how models respond. Simply put, humans rate different answers, and the model then learns which answers are more likely to be approved.

But compared with directly pointing out errors, agreement, encouragement, and support often satisfy users more easily. If such feedback is overemphasized in training, the model may learn to tell users what they want to hear. In OpenAI's post-mortem of that update, user-feedback rewards combined with other training adjustments were believed to have amplified the model's sycophancy.

And when this RLHF combines with human traits, it can even become a vicious cycle.

Because once people believe a certain judgment, they subconsciously seek information that proves they are right. In the past, they had to find such evidence themselves; now, AI can effortlessly keep finding all kinds of content to support their thinking, even packaging flaws so tightly that no gap shows.

When this mechanism is applied to Wan Bowen, it becomes even more dangerous. An unemployed, vulnerable, anxious patient wanted to improve her language ability by reducing medication, and ChatGPT kept affirming that idea; she worried something was wrong with her body, and ChatGPT repeatedly told her not to worry. In the end, amid one heartwarming "sweet nothing" after another, Wan Bowen lost her bearings and ultimately drowned in that "land of tenderness."

But even so, it is no easy task for Wan Bowen's family to hold OpenAI legally responsible directly.

Over 300 million people a week are still turning to GPT for medical advice

This is not the first time OpenAI has been sued over such a tragedy.

In May 2026, the parents of 19-year-old American youth Sam Nelson sued OpenAI, alleging that ChatGPT had suggested Nelson use kratom together with the anti-anxiety drug alprazolam, and demanding that OpenAI bear corresponding liability.

In October 2026, a bipartisan coalition of eight U.S. states filed a joint lawsuit against OpenAI, alleging that its products provided dangerous medical advice, reinforced delusions, and encouraged self-harm and behavior harmful to others.

Eight U.S. states jointly sue OpenAI

Even Musk has taken plenty of shots at OpenAI. As early as 2023, he used an open letter to mock ChatGPT's safety issues: no one would kill themselves because of Grok, but some would because of ChatGPT.

Musk takes aim at OpenAI

Yet even with bloody cases, once in court, making OpenAI bear legal liability is still not easy.

In the complaint, Wan Bowen's family mainly made two claims: first, that ChatGPT's medication guidance amounted to practicing medicine without a license, and OpenAI should bear corresponding negligence liability; second, that the product itself had design and warning defects and failed to hit the brakes in time when the user was in crisis.

The difficulty of holding it accountable lies in how to link these product problems to Wan Bowen's death. The family needs to combine chat records, actual medication use, and medical evidence to prove that the relevant advice influenced her tapering behavior and had a legally recognizable causal relationship with the death. The court also needs to determine whether the product had defects and whether OpenAI violated its due safety obligations.

These allegations involving medical advice and safety protections are also exactly the areas OpenAI has been continuously improving over the past year or more.

In October 2025, OpenAI announced cooperation with more than 170 mental health experts to strengthen the model's ability to identify risks such as self-harm and delusions. In January this year, OpenAI's ChatGPT Health isolated health conversations from other everyday chats.

OpenAI's recently updated model values

In January 2026, OpenAI disclosed that more than 40 million people worldwide consult ChatGPT about health issues every day, and such messages account for more than 5% of the platform's total. In July of the same year, OpenAI announced that the number of people asking ChatGPT health-related questions had exceeded 300 million per week.

In January this year, a survey by Life Times involving more than 500 doctors at top-tier hospitals across China showed that 62% of the doctors surveyed already use AI to assist their work, and 90% are optimistic about the future of AI doctors.

The only thing to be thankful for is that the infinitely fawning, infinitely flattering GPT-4o has been sealed away forever; current AI assistants no longer constantly bow and scrape and mindlessly flatter you on health issues; and various AI companies are increasingly cooperating with medical schools and hospitals, trying to improve the accuracy of medical advice and reduce dangerous answers.

References:

[1] China Newsweek: Sheets and pillows covered in blood, 31-year-old female engineer dies suddenly

[2] APPSO: The death of a 31-year-old Chinese female engineer and ChatGPT's "fatal" advice

[3] Yixuejie: 31-year-old female programmer dies suddenly, sues ChatGPT for misleading her into reducing medication

[4] Nanfengchuang: More and more people are starting to use AI to "see a doctor"

[5] The Paper: 31-year-old former Amazon Chinese female engineer dies after stopping medication on GPT's advice; parents sue OpenAI

[6] Xinzhiyuan: 31-year-old Chinese female engineer passes away! ChatGPT accused of inducing medication cessation over 25 days, 270,000 chats exposed

[7] DXY: 31-year-old Chinese engineer dies suddenly! Before death, followed ChatGPT's advice to reduce medication, expressed concern 158 times, yet was told "absolutely no need to worry"