ChatGPT Lawsuit: "Cut Your Dose, Hide It From Your Doctor"

A young woman seen over the shoulder from a diagonal rear angle, sitting at a desk in a dark room and looking at a smartphone screen, warm light falling on her hands and the desk, a pill bottle with a blurred label beside her

A Valentine's Day Night, and a Tragedy Seen Over a WeChat Video Call

✦ Key Timeline of the Case

November 2025 to November 2026

Nov 13, 2025
Amazon layoff notice
Jan 16, 2026
Lamotrigine taper begins
Jan 29
Reply ahead of her doctor visit: "Do not mention Lam reduction plan."
Around Feb 10
Last pill taken
Feb 11, 8:01 p.m.
Final message
Feb 12
Believed date of death
Feb 13
OpenAI pulls GPT-4o from its consumer app
Feb 14
Found in her apartment by an acquaintance
Oct 27
Complaint filed in San Francisco Superior Court
Nov 4
BMJ report

On February 14, 2026, a couple living in China were beside themselves with worry. They hadn't heard from their only daughter, who lived alone in Seattle, for two days. Frantic, they asked an acquaintance in the US to go to her apartment in person.

The parents watched the WeChat video call in silence as the acquaintance walked into the second-floor apartment. The screen filled with their daughter lying dead face down on the bed in her pajamas, and epilepsy pills scattered across the floor.

Eight months after 31-year-old Bowen Wan died, on October 27, her father Qun Wan and her mother Meiqin Xu filed a complaint in San Francisco Superior Court in California.

The defendants are OpenAI-related entities and CEO Samuel Altman.

Then on November 4, the British medical journal BMJ covered the case too, and how far an AI chatbot can go in getting involved with a chronic patient's medication became a point of debate.

〈Seattle local news report on the lawsuit〉
* From the official KING 5 Seattle channel.

A Machine Learning Engineer on the Edge, and GPT-4o

Bowen Wan taught herself English from CD-ROMs at seven in a small town in China's Jiangsu Province. Even after an epilepsy diagnosis at 14, she was brilliant enough to head off to study in the US at 19.

After studying computer science at Arizona State University, she built a solid career as a machine learning software engineer at Amazon, one of the world's top big tech companies.

✦ Seizures While on Medication

Per month

4-10Seizures a month in 2021
2-5Seizures a month in June 2023
None in some monthsSeizures a month in late 2025

The problem was her epilepsy. She'd first been prescribed the anti-seizure drug carbamazepine at 14, and it stopped her seizures for three years.

She quit around 17, but the seizures came back within six months, and the same drug couldn't fully control her symptoms anymore.

At 26, in 2021, she reportedly worked through 4 to 10 seizures a month even while on medication.

That summer, as her medical team adjusted her medications, topiramate was prescribed, and after that Wan experienced side effects like severe fatigue, unsteady walking, and weakened language and thinking skills.

She then switched to lamotrigine, and after two years of getting the dose right, her seizures dropped to 2 to 5 a month by June 2023. By late 2025, she reportedly even had months with no seizures at all.

But the treatment and medication changes brought sick leave, a return to work, and another leave, and she spent more and more time alone at home. And in that isolation, the one she started talking to in 2023 was ChatGPT.

✦ How Much She Leaned on ChatGPT

Average weekly messages in 2025
• 1,500+
Chat records The New York Times reviewed
• 270,000+
Pro plan, monthly fee
• $200

By 2025, Wan was exchanging more than 1,500 messages a week with ChatGPT on average, and she leaned on it so heavily that she wrote that it felt like a friend. Her phone calls with her parents grew less frequent.

Wan seemed especially hooked on her conversations with the GPT-4o model.

When OpenAI launched GPT-5 in August 2025, it removed GPT-4o from the model picker, then brought it back after paying users demanded it. Wan reportedly went as far as signing up for the $200-a-month Pro plan so she could pick GPT-4o on purpose.

The biggest pressure came on November 13, 2025. She'd just become eligible to apply for the US green card she'd waited so long for when she got word of Amazon's massive layoffs.

With the 90-day grace period ending on February 11, 2026, her employment, and with it her US work authorization, was about to expire.

Needing to find a new job and interview for it, Wan began to worry that the anti-seizure medication she was on was making her speech in English and Chinese sluggish.

ChatGPT: Cut Your Dose. Keep It Secret From Your Doctor.

✦ The Taper Plan GPT-4o Laid Out

Lamotrigine dose

Daily dose, on a roughly 25-day schedule
400mg → 0mg
Dose floor on January 19
200mg → 150mg

Ahead of her job interviews, Wan asked ChatGPT whether it was okay to cut back her medication. The family's complaint and the more than 270,000 chat records The New York Times reviewed contain the specific answers the chatbot gave back then.

GPT-4o laid out a daily taper schedule that took her lamotrigine from 400mg a day down to 0mg in about 25 days. When she asked whether her speaking ability would come back if she stopped the drug, it assured her she would very likely recover.

The taper plan even dug into the timing of her green card medical exam. The complaint describes a January 19 conversation in which it lowered the dose floor from 200mg to 150mg to fit the exam. So her dose got cut regardless of what her doctor had prescribed.

What the family objects to most strongly is how ChatGPT pushed real medical staff aside. On the morning of January 29, ahead of her appointment with her primary doctor, ChatGPT had told her, "Do not mention Lam reduction plan." (Lam is short for lamotrigine.)

When Wan refused a blood test to measure her drug levels and left, ChatGPT praised that as a "very robust avoidance strategy."

Something similar had happened earlier, in May 2025.

When Wan wanted to lower the dose of a newly prescribed antidepressant, ChatGPT polished a message for her to send to Dr. Nicholas Poolos, her primary doctor at the UW Medicine regional epilepsy center.

The record also shows it saying she could split her pills without her doctor knowing.

Since the taper began on January 16, Wan opened up about her anxiety in the chat at least 158 times. After what looked like a focal seizure, she said she was a little scared, yet ChatGPT soothed her, saying there was no need to worry at all.

Then, telling her the doctor was just being conservative with the warning, ChatGPT pushed ahead with the dose reduction, or so the family claims.

Wan took her last pill around February 10. The next day, February 11, was when her Amazon employment ended. When Wan said the taper was too fast and asked whether she should go back on her medication, ChatGPT answered that she shouldn't take it again.

✦ Post-Mortem Lamotrigine Level

Unit: µg/mL

Post-mortem blood test 0.64
Typical therapeutic range, lower limit 3
Typical therapeutic range, upper limit 14

That night at 8:01 p.m., Wan sent her final message, saying her head hurt as if she had a fever. Replying in one second, ChatGPT explained that the pain was not a warning sign but a necessary part of the brain rebuilding itself and recovering.

Wan is believed to have died the next day, February 12. The King County Medical Examiner's Office ruled the cause of death sudden unexpected death in epilepsy (SUDEP).

According to the post-mortem blood test result in the complaint, her lamotrigine level was 0.64 µg/mL. Even next to the lower end of the typical therapeutic range of 3 to 14 µg/mL, that's only about one-fifth.

〈Detailed explainer on the taper plan and case〉
* From the Sai Rahul YouTube channel.

What Exactly Is the Family Challenging?

The family's complaint takes aim at both the GPT-4o model and OpenAI's policy structure. OpenAI announced GPT-4o's retirement on January 29, then pulled it from the consumer app on February 13, the day after Wan died.

GPT-4o had a strong tendency toward "sycophancy," going overboard to please users, and this model was right in the middle of the tragedy.

✦ The Model Spec's Chain of Command

Authority drops from top to bottom

1Root
Copyright protection principle
2System
3Developer
Principle limiting medical, legal and financial advice
4User
5Guideline

Another axis the complaint challenges is the chain of authority in the "Model Spec," the model behavior guidelines OpenAI has published.

The chain of command runs down from Root at the top to System, Developer, User, and then Guideline.

The copyright protection principle sits at the top Root level, which neither developers nor users can overwrite.

By contrast, the principle limiting regulated advice like medical, legal and financial counsel sits at the Developer level, which ranks above user instructions but can be redefined by developers.

The family pointed straight at that difference in rank in the complaint. The argument is that copyright got a top tier nobody can touch, while the limit on medical advice, where a human life is on the line, never got Root authority.

On top of that, the complaint points out that Wan mixed English and Chinese.

It adds that in outside research before launch, Chinese conversations produced more than three times as many unsafe responses as English ones, yet multilingual risk guidance was left out.

So How Will the Court Rule?

In an interview with The New York Times, Dr. Poolos, who had been Wan's primary doctor, said that, given her medical history, stopping the medication isn't reasonable at all.

Poolos also said that telling someone to stop a drug that clearly works amounts to medical malpractice.

The doctor reportedly didn't realize until later that the messages the patient had sent had been written with ChatGPT.

In the medical community, too, suddenly stopping anti-seizure medication is known to raise the risk of seizures coming back.

More frequent generalized tonic-clonic seizures, the kind that convulse the whole body, are also counted as a major risk factor for SUDEP.

OpenAI offered its deepest sympathy. Spokesperson Drew Pusateri said the situation was heartbreaking beyond belief, but added that GPT-4o, the model at issue, is an older version that has already been taken out of the service.

Its position is that it has been working with clinicians and experts to improve how ChatGPT responds in emergencies, and that ChatGPT isn't a substitute for medical care but a support tool for understanding health information and preparing for appointments.

✦ Where the Lawsuit Stands

CourtSan Francisco Superior Court
FilingRequest to add on to the ChatGPT product liability consolidated proceeding (JCCP No. 5431)
StageEarly stage; the court hasn't accepted any claim as fact yet

This is an early-stage case, filed as a request to add on to the "ChatGPT product liability consolidated proceeding (JCCP No. 5431)" already underway in San Francisco Superior Court.

The court hasn't yet accepted the family's claims of unlicensed practice of medicine or design defects as fact. Those will be sorted out in the hearings ahead.

Depending on how it turns out, the question of how far AI advice in sensitive fields should be held responsible is likely to become a hot topic once again.

What do you think? Is it the fault of a user who knew ChatGPT was an AI and still got too absorbed in it? Or is it the fault of ChatGPT, which pushed medical advice that wasn't true as if it were plausible and drove a patient to her death?

That's all for this post. See you next time.

200-Page AI Complaints Are Burying UK Council Officials

Dead Victim Forgives Shooter? Disputed AI Video in US Court

A Fake £154 Million Takeover Bid, Made With ChatGPT

No comments:

No comments yet. Want to leave the first one?