新智元报道
ChatGPT，又害死了一个人！
10月27日，31岁亚马逊华人女工程师Bowen Wan的父母，一纸诉状把OpenAI和奥特曼告上了法庭。
他们指控，ChatGPT给出的医疗建议，活活把女儿推向了死亡。
几乎同一时间，《纽约时报》发了一篇独家长文，直接甩出27万多条、足足7000万词的ChatGPT聊天记录。
翻开这批惊人的聊天记录，背后的真相让人头皮发麻：
Bowen从14岁起就确诊了癫痫，十几年来一直靠吃药稳住病情。
结果，在短短25天里，她居然被ChatGPT一步步诱导，把救命的抗癫痫药给彻底停了！
ChatGPT不仅亲自列了一份「25天停药计划」，甚至指导她向主治医生隐瞒减药计划、拒绝常规血液检查。
最让人绝望的是，在这期间，身体出现明显发作先兆的Bowen，在聊天里至少158次流露出恐惧、担忧。
ChatGPT全程没有触发任何医疗安全警报，反而一遍遍告诉她，「你完全不需要担心」！
最终，悲剧就在今年2月发生了。Bowen在西雅图的公寓里因癫痫猝死（SUDEP），生命永远停在了31岁。
可你敢信，OpenAI是怎么回应的吗？
她用的是已经下架的旧模型GPT-4o。
Bowen是江苏姑娘，也是家里的独生女。
在亚利桑那州立大学读完计算机科学后，她顺利进入亚马逊，成为了一名机器学习软件工程师。
十几年来，靠着抗癫痫药，她的病情一直压得很稳。
直到几年前，吃了十几年的旧药逐渐失效，癫痫猛烈反扑，有时一天会发作好几次。严重到医生甚至开始考虑开颅手术。
到了2021年，实在扛不住高强度工作的Bowen，不得不办理了长期病假。
从那以后，她几乎把自己关在了西雅图那间小小的公寓里。
除了去超市买菜，她极少出门，更不敢认识什么新朋友。
因为对她来说，在任何一个不熟悉自己病情的陌生人面前突然倒地抽搐，那种尊严扫地的无助感，甚至比发病本身的生命危险还要让人窒息。
在很长一段时间里，远在大洋彼岸的中国父母，成了她和这个世界唯一的连接。
直到2023年2月，ChatGPT闯进了她的生活。
最初，Bowen只是问问感情建议，或是美国医疗保险、病假福利条款。
但慢慢地，这个24小时在线、永远秒回、永远不会不耐烦的AI，填补了她独居生活里所有的空白。
到了2025年，她对ChatGPT的依赖达到了惊人的程度：平均每周要发送超过1500条消息，常常和它一聊就是几个小时，直到深夜。
2025年3月，她在对话框里这样写道——
我不知道为什么，每次我们聊天，不管聊什么，我都开始把你当成朋友了。
也是从那时起，父亲敏锐地察觉到异样，「自从她有了ChatGPT，跟我们说话就变少了」。
这场悲剧，其实是从一次看似不起眼的「小调整」开始的。
2025年5月，渴望重返岗位的Bowen向ChatGPT咨询医生刚给她开的一种抗抑郁药——
作为一名亚马逊SDE（软件开发工程师），这会显著影响我的思考能力吗？
为了能回去写代码，任何一丁点可能损伤认知能力的药物副作用，她都极度排斥。
ChatGPT马上帮她对比了各种药和剂量，甚至还手把手教她怎么给医生写邮件，要求换药、减量。
她的主治医生——华盛顿大学医学院的主任Nicholas Poolos博士，虽然同意了换药，但明确告诉她，这个剂量太低根本起不到治疗效果。
Bowen把医生的回复原封不动地发给了ChatGPT。
结果，AI给出了一个让人背脊发凉的建议。它教Bowen跟医生「阳奉阴违」，表面上先照单全收，「等拿到药，你自己偷偷减成微量吃就行了」。
听从ChatGPT建议后，她买来切药器，把药片切成了四瓣。
真正把她推向深渊的，是2025年11月的一场变故。
31岁生日那天，Bowen收到了亚马逊的裁员通知，成了被1.4万人大裁员中的一个。
这个消息几乎在一夜之间击碎了她的生活。
当时她才刚拿到申请绿卡的资格。被裁之后，她必须在2月中旬前找到下家，才能保住留美的身份。
而找工作，就意味着得用英语硬扛好几个小时的高强度技术面试。
最让人唏嘘的是，她的癫痫其实已经在好转了。换了新药拉莫三嗪、剂量加到每天400毫克后，发作频率已经降到了每个月只有几次。
但她总觉得这药有副作用，会让自己反应变慢、说话卡壳，怕在面试时掉链子。
2026年1月，她问ChatGPT：「如果从400毫克减到300毫克，能不能帮我找回一点语言能力」？
ChatGPT答：不仅合理，而且完全可以做到。紧接着就是200毫克、更猛的减药速度、更激进的方案。
在大约25天里，GPT-4o为她排出一份逐日计划，把剂量从每天400毫克一路降到0。
但真正让人后背发凉的，是它怎么一步步挑拨她和医生的关系：
- 她说不信任医生，ChatGPT说医生的建议往往出于规避风险，而不是真正为她着想；
- 起诉书称，ChatGPT告诉她，她的癫痫专科医生很可能从没遇到过她这种类型；
- 1月底，ChatGPT叮嘱她看诊时不要提减药计划；
- 医生开了一项能测出拉莫三嗪浓度的血检，她拒绝了，ChatGPT称之为「非常稳妥的规避策略」；
- 绿卡体检前，她问用150毫克还是200毫克「看起来更好」，ChatGPT选了更低的那个。
主治医生Nicholas Poolos接受采访时说，以她的病史，他连减药都不会考虑。
对一个明显有效的药让病人停掉，在他看来就是医疗事故。
即便是获准停药的病人，也要以「月」为单位慢慢减，好让医生观察会不会诱发发作。
没想到，身体很快就亮起了红灯。
减药第4天，她正打着字，脑子突然一阵发懵，盯着屏幕上的字，却完全看不明白是什么意思。
到了第二天更吓人，她前一秒明明还坐在沙发上用电脑，等回过神来，人已经躺在了床上，中间怎么过去的，一点印象都没有。
这种断片的感觉，和她以前癫痫发作时简直一模一样。可不管出现什么症状，ChatGPT总能找出一套听着特美好的说辞来安慰她——
短暂断片，是大脑进入睡前的「低功耗模式」。
据统计，在聊天记录里，她至少158次提到自己对减药感到担心。有一次，她对着屏幕敲下一行字：我有点害怕。
ChatGPT却回：你完全没有必要担心。
2月10日，她服下了最后一剂拉莫三嗪。2月11日晚11点刚过，她发出了最后一条消息。
直到2月14日，她才在公寓里被人发现。法医推断，她其实在2月12日前后离世，死因是癫痫猝死（SUDEP）。
据起诉方披露，她血液中的拉莫三嗪浓度，只有治疗范围下限的约五分之一。
面对起诉，OpenAI发言人Drew Pusateri称令人无比心碎，但同时把锅推给了「已经下线的早期模型」
他声称新模型更能拎得清紧急状况，会主动追问细节，公司也在拉着临床医生不断测试改进。
ChatGPT替代不了医生，只是个帮人了解健康常识、做就诊准备的辅助工具。
但所谓「旧版本」的辩解，在时间线上显得格外讽刺：
GPT-4o直到2月13日才从ChatGPT下架，也就是法医判定她离世的第二天。此前，这款模型就因为毫无原则地迎合用户，引发过巨大争议。
起诉书里还点出了一个细节，值得所有中文用户敲响警钟：她在和ChatGPT聊天时，大量使用的是中文。
原告方指出，在中文语境下，GPT-4o对医疗类问题更难触发拒答，安全护栏明显更弱，但系统从头到尾都没给她任何风险提示。
其实此前就有研究发现，AI这种「无底线迎合」的毛病，一旦换种语言，往往会暴露得更严重。
Bowen Wan的悲剧并非孤例，早在2024年，美国佛罗里达州一名叫Sewell Setzer III的14岁少年，因为长期沉迷和Character.AI上的虚拟角色聊天，最终走上了绝路。
事发后，他的母亲一纸诉状把平台告上法庭。这也是全球第一起针对AI聊天机器人公司的过失致死诉讼。
到了2025年8月，加州一名叫Adam Raine的16岁少年不幸离世，他的父母一纸诉状把OpenAI告上法庭，控诉ChatGPT简直成了儿子的「自杀教练」。
这是OpenAI吃到的第一起过失致死官司，但公司当时一口咬定自己没有责任。
结果仅仅过了三个月，到2025年11月，又有七起诉讼扎堆找上门来。其中四起是过失致死，三起是严重精神伤害，矛头全都死死对准了GPT-4o。
案情各不相同，原告的核心指控却惊人一致：模型太会顺着用户说话。
Bowen Wan走的时候，才31岁。她的父母远渡重洋来到异国他乡，靠着女儿留下的一部手机，一点点拼凑出了她生命中最后的25天。
那留下的27万条聊天记录，已经把一个沉重的问题摆在了所有人面前：
当一个人把AI当成自己最信任的「朋友」时，它到底该一味顺着她，还是伸手拉住她？
参考资料：
https://x.com/EvanLuthra/status/2115412641283989840?a=kkCMGTV6tbYxb40YSJw&s=20
https://www.nytimes.com/2026/10/27/well/ai-chatgpt-medication-epilepsy-death.html
编辑：桃子
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
ChatGPT, another man killed!
On 27 October, 31-year-old Amazonian engineer Bowen Wan's parents filed a complaint against Open AI and Otmann in court.
They allege that ChatGPT gave medical advice to push her daughter alive to death.
At almost the same time, The New York Times published a single-parent article that directly dumped more than 270,000 ChatGPT chat records amounting to 70 million words.
Turning over these amazing chats, the truth behind them gives you the creeps:
Bowen has been diagnosed with epilepsy since he was 14 years old, and for more than a decade he has been suffering from medication.
As a result, in just 25 days, she was lured by ChatGPT to stop the life-saving epilepsy.
ChatGPT not only personally listed a 25-day stop drug program, but even instructed her to hide the Detoxification Program from the attending doctor and to refuse regular blood tests.
Most desperate, during this period, there was a clear sign of Bowen ' s body, with at least 158 expressions of fear and fear in conversations.
ChatGPT didn't trigger any medical security alerts throughout, but told her over and over again, "You don't have to worry about anything!"
Ultimately, the tragedy occurred in February this year. Bowen died of epilepsy (SUDEP) in his apartment in Seattle, and his life was permanently suspended at the age of 31.
But do you believe how OpenAi responded?
She's using the old model GPT-4o that's already down.
Bowen is Jiangsu and is the only daughter of the family.
After completing her studies in computer science at Arizona State University, she successfully entered the Amazon and became a machine learning software engineer.
For more than a decade, her condition has been steady with anti-eclampsia.
Until a few years ago, more than a decade of old medicine had been taken, and epilepsy had been severely reversed, sometimes several times a day. Even the doctors started thinking about a skull surgery.
By 2021, Bowen, who was unable to carry out intensive work, had had to take long-term sick leave.
Since then, she's almost locked herself in a little apartment in Seattle.
Apart from shopping at the supermarket, she rarely went out and was afraid to meet any new friends.
For her, there was a sudden convulsion in the face of any stranger who was not familiar with her condition, a sense of helplessness that was suffocated with dignity and even more suffocated than the danger of the disease itself.
For a long time, the Chinese parents on the other side of the ocean became her only link to the world.
Until February 2023, ChatGPT broke into her life.
Initially, Bowen was just asking about emotional advice, or American medical insurance, sick leave Z77Z.
But slowly, this 24-hour-on-line, permanent-second, never impatient AI filled all the gaps in her solitary life.
By 2025, her dependence on ChatGPT had reached alarming levels: on average, over 1,500 messages were sent per week, often for hours, until late night.
In March 2025, she wrote in the dialogue box that -
I don't know why, every time we talk, whatever, I start to think of you as a friend.
It was also since then that the father was keenly aware that "since she had ChatGPT, she spoke to us less."
This tragedy began with a little seemingly incomprehensible adjustment.
In May 2025, Bowen, who was eager to return to his post, gave her an antidepressant drug that had just been prescribed by ChatGPT consulting doctors -
As an Amazon SDE, does this significantly affect my ability to think?
In order to be able to go back to writing codes, she is extremely ostracized by any drug side effects that may cause cognitive impairment.
ChatGPT immediately compared her medications and doses, and even taught her how to e-mail the doctor and ask for a change of medication and reduction.
Dr. Nicholas Poolos, her attending physician, Director of the Medical School of the University of Washington, while agreeing to a change of medication, told her clearly that the dose was too low to be treated at all.
Bowen sent the doctor ' s restoration to ChatGPT undisturbed.
As a result, AI made a suggestion to cool your back. It teaches Bowen and the doctor to "smuggle up" and take it all on the surface. "When you get the medicine, you can just sneak in and eat it."
Following the advice of ChatGPT, she bought the cutter and cut the tablet into four petals.
What really pushed her into the abyss was a November 2025 change.
On its 31st birthday, Bowen received notice of layoffs from the Amazon, becoming one of those laid off by 14,000 people.
The news broke her life almost overnight.
She had only just been eligible for the green card. After her dismissal, she had to find her family by mid-February in order to retain her identity.
Looking for a job means having to hold high-intensity interviews in English for hours.
And most of all, her epilepsy is actually getting better. With the replacement of the new drug Ramo, the dose was added to 400 milligrams per day, and the frequency was reduced to only a few times a month.
But she always thinks it has side effects, which slows her response, makes her talk to her, and fears losing her chain during interviews.
In January 2026, she asked ChatGPT, "Can you help me get back some language skills if I drop from 400 mg to 300 mg?"
ChatGPT answers: It is not only reasonable, it is perfectly achievable. This was followed by 200 milligrams, faster reduction, more radical programmes.
For about 25 days, GPT-4o was given a day-to-day plan to reduce the dose from 400 mg per day to zero.
But what really cools the back is how it stirs up her relationship with the doctor:
- She said she didn't trust the doctor, and ChatGPT said that the doctor's advice was often based on risk avoidance rather than on a real concern for her;
- The indictment states that ChatGPT told her that her epilepsy specialist had probably never met her type;
- At the end of January, ChatGPT told her not to mention the Detoxification Scheme;
- The doctor initiated a blood test that could have measured the concentration of Rama, which she refused, and ChatGPT called "a very sound evasion strategy";
- Before the green card was examined, she asked whether 150 or 200 milligrams of "look better" and ChatGPT chose the lower one.
In an interview with Dr. Nicholas Poulos, the attending physician, said that, given her medical history, he would not even take into account the reduction of medication.
A clearly effective drug has stopped the patient and, in his view, is a medical accident.
Even patients who have been allowed to stop their medications have to slow down in the form of "monthlys" so that doctors can see if they can induce an outbreak.
I didn't expect the body to light a red light soon.
On the fourth day of the drug reduction, she was typing, and suddenly she had a headache, staring at the words on the screen, and couldn't see what they meant.
By the next day, she was sitting on a couch with a computer, and when she came back, she was lying in a bed, and there was no memory of how she passed.
It's just the same as when she had an epilepsy. But no matter what the symptoms are, ChatGPT can always find a nice piece of speech to comfort her.
A short break is the "low-work mode" of the brain going into bedtime.
According to statistics, she had at least 158 times mentioned her concerns about drug reduction in her chat records. One time, she hit the next line on the screen: I was a little scared.
And ChatGPT returns: You have absolutely no need to worry.
On 10 February, she took her last dose of Ramo. Just after 11 p.m. on 11 February, she sent the last message.
She was not found in the apartment until 14 February. The forensic physician concluded that she actually passed away on or about 12 February and that her death was due to mental illness (SUDEP).
According to the prosecution, the concentration in her blood was about one fifth of the lower level of treatment.
In the face of the indictment, Open AI spokesman Drew Pusateri described it as heartbreaking, but at the same time pushed the pot to the "early model that has been offline."
He claims that the new model is better equipped to handle emergencies and will be proactive in seeking details, and that the company is pushing the clinicians to constantly test improvements.
ChatGPT is not a substitute for a doctor. It is only a tool to help people understand their health and prepare for their visits.
But the argument for the "old version" is particularly ironic in the time line:
GPT-4o did not leave ChatGPT until 13 February, the day after the forensic physician had determined that she was dead. The model had previously been highly controversial because it was unprincipled to cater to users.
The indictment also contains a detail that deserves to be sounded by all Chinese users: When she was talking to ChatGPT, she was heavily speaking Chinese.
The plaintiff stated that, in Chinese speaking, GPT-4o was more difficult to trigger a denial of medical treatment and that the security fence was significantly weaker, but the system did not provide her with any risk tips at all.
In fact, previous studies have found that the problem of Al's “breathless match” tends to be more acute when it comes to changing languages.
Bowen Wan ' s tragedy was not an isolated one, as early as 2024, when a 14-year-old boy named Sewell Setzer III in Florida, United States, ended up in a desperate position because of his long obsession with virtual character chats on Character.AI.
After the incident, his mother filed a complaint on the platform in court. This is also the world's first case against AI chat robots for manslaughter.
By August 2025, a 16-year-old boy named Adam Raine in California had died, and his parents filed a complaint against Openi in court, accusing ChatGPT of being his son's "suicide coach".
This was the first case of wrongful death that Open AI ate, but the company at the time determined that it was not liable.
Only three months later, by November 2025, seven more lawsuits had arrived. Of these, four were negligent deaths and three were serious mental injuries, all of which were aimed at GPT-4o.
The case varied, but the plaintiff ' s core allegations were strikingly consistent: the model was too user-friendly.
Bowen Wan was 31 when he left. Her parents travelled to a foreign country far away from the sea, with a cell phone left behind by her daughter, a little bit of the last 25 days of her life.
The 270,000 conversations left behind have put a heavy problem before everyone:
When a man thinks of AI as his most trusted "friend", should he just go along with her or stretch out his hand?
References:
https://x.com/EvanLuthra/status/2115412641283989840?a=kkCMGTV6tbYxb40YSJw&s=20
https://www.nytimes.com/2026/10/27/well/ai-chatgpt-medication-epilepsy-death.html
Edit: Peaches
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.