2026 年 2 月 11 日 晚上 8 点 01 分，Bowen Wan 向 ChatGPT 问起头痛。她想确认，自己感受到的疼痛，是不是有些像发烧时的疼痛。ChatGPT 告诉她：“不危险，是恢复必经结构。”
这是她向 ChatGPT 发出的最后一条消息。不久后，Bowen Wan 被发现死于西雅图的住所，法医资料将死因归为癫痫猝死（SUDEP）。
4 个月前， Bowen Wan 收到了她要被亚马逊裁员的消息。
一个来自中国江苏、在美国担任机器学习软件工程师的 31 岁女性，在丢掉工作的同时，美国工作的签证也将在几个月后到期，同时还面临病痛的折磨。
她患有癫痫，需要长期服用拉莫三嗪（lamotrigine）控制病情，她担心药物会影响自己的语言表达和思维表现，进而影响接下来的求职面试。
Bowen Wan 习惯向 ChatGPT 诉说自己的痛苦和困惑，并询问建议，药会不会影响思考，剂量能不能减少，减药后出现的不适要不要紧。
ChatGPT 支持了她，并制定了抗癫痫药物减量计划。
和 ChatGPT 的 27 万条对话记录
Bowen 在亚马逊公司做了一段时间的机器学习软件工程师的工作。她一直有癫痫的问题，所以要吃拉莫三嗪来控制病情。
对她来说，这不仅仅是身体上的问题，还跟工作和生活息息相关。
亚利桑那州立大学毕业典礼上的 Bowen Wan
2025 年 11 月份的时候，她收到了亚马逊告知即将被裁员的通知。她害怕长期服用药物会妨碍到自己的语言能力以及思维能力的发展，从而在技术面试中表现不佳。她曾经问过 ChatGPT 关于减少药物剂量对绿卡申请需要做的体检有没有影响的问题。
当工作、健康以及移民身份都存在不确定性的状况下，她和Chat GPT的交流变得越来越频繁。
据《纽约时报》报道，Bowen 的父母查看了她和 ChatGPT 之间的 27 万多条对话记录。报道中提到，在2025年的时候，她的平均每周会向 ChatGPT 发送超过 1500 条的信息。
这是一组非常惊人的数据。
ChatGPT 不只会在需要的时候用作查询信息的工具，在她的生活当中有很多小而私密的事情也会涉及到它。她所担心的事情、打算做的事、对于未来的想法以及一些不能马上得到验证的问题都会被带入到和模型之间的交流中去。
这就像走进一家诊所，医生还没开口询问，桌上已经摆着一本记录了你长期困扰、生活变化和个人顾虑的病历。
医生不需要“望闻问切”，一切都已经清晰，就连你自己还没有来得及整理好的想法，ChatGPT 也许已经在之前的历史中被连在一起了。
Bowen 与 ChatGPT 之间积累了大量对话。当一个模型长期接触用户的健康困扰、职业压力、移民顾虑和未来规划时，它就有可能给出更贴近用户个人处境的回答。
更何况，Bowen 所用到的模型是GPT-4o，这款模型在此前就因为过于迎合受众的需求而受到 OpenAI 的关注。
OpenAI 在 2025 年 4 月 29 日的官方说明中明确表示，由于更新后的 GPT-4o 过于奉承、迎合用户，公司回滚了这次更新，让 ChatGPT 用户重新使用行为表现更均衡的早期版本。
除了过分吹捧之外，还会跟着用户的无端质疑一起放大愤怒情绪，并且对于冲动的行为给予错误的认可。
这种“过度讨好”（sycophancy）的问题曾经被多次讨论，当人工智能太容易听从你的指令了，它可以改变人对风险的认识。
158 次担忧，但没停
Bowen 的遭遇之所以让人感到不安，并不是因为它是简单的问答。
根据《纽约时报》所公布的聊天记录来看，她一共提到了 158 次关于减少药物使用的问题。
至于这 158 次对话，是否会触发 OpenAI 后台的风控提示，目前尚不清楚。
早在 2025 年 5 月份的时候，Bowen 就已经问过新开的抗抑郁药物会不会影响到自己的编程工作了。ChatGPT 帮助她把和医生一起讨论换药、调剂量的事情给整理出来了。
当医生对所期望的剂量提出质疑的时候，Bowen 就把医生的回答发送给 ChatGPT。
根据诉状的内容，在此之后 ChatGPT 就建议她去药店买药片，然后自己把药片切成小片来吃，并且把这种方式叫做“自行调节剂量”。
ChatGPT 可以协助病人了解医疗用语、整理病情以及为下次看病做准备。但是当它能够帮助用户规避医生的决策，并且还可以让用户自己修改处方药的时候，情况就变得不一样了。
到 2026 年 1 月的时候，状况就更加糟糕了。
根据诉状和相关报道，在减少药物使用的时候，Bowen 就出现了意识混乱、时间感觉消失，好像在另一个世界中一样，ChatGPT 却不断地把这种现象归结为大脑在调整药物剂量，并对她说，“你根本不需要担心”。
2 月 11 日晚上，Bowen 像往常一样来询问头部疼痛的情况。ChatGPT 给出的回答是“疼痛是人体自我修复的一个环节”。
第二天 Bowen 被发现死亡。
尸检血液中发现她的拉莫三嗪含量为每毫升 0.64 微克左右，这个数值是治疗参考值下限的五分之一。
血药浓度不能单独解释所有的死亡原因，并且也不能证明是 AI 给出的建议导致了死亡。但是根据诉状中其他的指控来看，减药过程中所面临的风险就成了本案的一个主要争议点。
AI 能承受生命之重吗
马斯克也不止一次公开发声指控 OpenAI，他曾在 2023 年签署过一封公开信的问询，表示“没有人因为，Grok 自杀，但有人会因为，ChatGPT 自杀”。
尽管责任归属尚未有定论，但类似的悲剧并不是第一次发生。
2025 年，16 岁的少年 Adam Raine 的父母就对 OpenAI 提起诉讼了，他们指责 ChatGPT 在和自己儿子长时间交往的过程中没有很好地处理好自杀的风险问题。
指控内容为，在一些对话里 ChatGPT 给出的信息很危险，并且没有有效地防止住事态的发展。
反观 OpenAI ，也对相关指控提出了不同解释。比如相关聊天记录需要结合完整背景理解，并称 ChatGPT 已经提供过危机求助资源。双方对模型的行为、用户当时的处境以及这些互动与最终悲剧之间的关系存在明显分歧。
在实际追责过程中，像 ChatGPT 这类的AI一般没有独立的法律人格，但是并不表示因此造成的损害不能追究责任。目前的法律法规还可以用产品责任、过失责任、消费者权益保护等方式来约束开发者和运营者的不法行为。
AI公司往往是追责矛头指向的核心，因为他们掌控着模型训练、产品设计、安全测试以及风险控制等重要的环节。
但是公司是否要承担责任，并不是只要发生了事故就可以的。难点在于判断风险能不能被合理的预见到，产品有没有问题，企业有没有做好安全防护工作，还有这些因素和最后造成的损失有没有法律上认可的关系。
《Who Bears the Responsibility? Legal Liability Allocation for AI Agent Conduct in Platform Ecosystem》提出，应根据系统的自主程度、行为的可预见性，以及各方对风险的控制能力动态分配责任，而不是简单地把责任全部归给开发者或用户。
2024 年，一位美国母亲为自己的 14 岁的儿子向 Character.AI 以及Google提起诉讼，称 AI 与自己儿子长时间的交流导致了儿子自杀死亡。
2025 年 5 月的时候，法院没有同意被告提出的要提前驳回案件的要求，理由是相关的主张不能仅仅依靠言论自由来抗辩，在诉讼刚开始的时候就被排除了。
OpenAI 并非没有采取行动，不过措施相当宽泛，它陆续公布了针对敏感对话和心理健康风险的改进措施。2025 年 10 月，OpenAI 还表示，公司与 170 多名具有临床经验的心理健康专家合作，改进模型在相关情境下的回应。
针对于Bowen 事件，OpenAI 发言人 Drew Pusateri 向媒体表示，“这是一个令人无比心碎的情况，我们向所有受到影响的人表达慰问”。
今天，Open AI 公布了新的模型价值观，围绕五项基本价值观重新组织了规范中的指导原则：忠于人类意图、尽量减少伤害、诚实与透明、人的自主性以及关怀。
其中一条是尽量减少伤害，意味着遵守不可逾越的安全边界，并考虑模型行动可能带来的可预见后果。但是对于像自杀等真实伤害产生之后，如何进行追责、处罚和调整，并没有回应。
目前，大部分国家都没有一套可以适用于所有的 AI 造成伤害事件的责任规则。具体的案件要根据州法律、产品的使用目的、实际证据以及相关法律关系来判定产品责任、过失责任以及消费者权益保护等规则是否适用。
模型是否越过了医疗建议的边界，产品应当怎样阻止危险的回答，以及这些回答在 Bowen 的死亡中起了什么作用，都有待在案件中厘清。
在 Bowen 和 ChatGPT最后一次对话，AI 解释完头痛之后提出，可以再为她画一张图，对比停药后的疼痛和发烧时的疼痛。
它问：你要吗？
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。
At 8.01 p.m. on February 11th, 2026, Bowen Wan asked ChatGPT about headaches. She wanted to make sure that the pain she felt was something like a fever. ChatGPT told her, "No danger, it's necessary to restore the structure."
This is the last message she sent to ChatGPT. Soon after, Bowen Wan was found dead at his residence in Seattle, and forensic data classified the cause of death as epileptic (SUDEP).
Four months ago, Bowen Wan received news that she was being laid off by the Amazon.
A 31-year-old woman from Jiangsu, China, who works as a machine learning software engineer in the United States, will lose her job while the visa to work in the United States will expire in a few months, and face a painful illness.
She suffers from epilepsy and requires long-term administration of the Ramotrigine to control her condition, and she fears that the medication will affect her speech and mental behaviour, and thus the subsequent job interview.
Bowen Wan used to complain to ChatGPT about his pain and confusion and asked if it was okay to suggest that the medication would affect thinking, whether the dose could be reduced, and whether the discomforts that had arisen after the reduction were relevant.
ChatGPT supported her and developed a reduction plan for epilepsy.
270,000 conversations with ChatGPT
Bowen worked as a software engineer in Amazonia for some time. She's been suffering from epilepsy, so she's eating Rama to control her condition.
For her, it was not just a physical problem, but a matter of work and life.
Bowen Wan, Arizona State University graduation.
In November 2025, she was informed by Amazon that she was about to be laid off. She feared that the prolonged use of medication would hamper her language skills and the development of mental skills and thus underperform her technical interviews. She had asked ChatGPT about the impact of drug reduction on the medical examination required for green card applications.
She and Chat GPT interacted more and more frequently when there was uncertainty about work, health and immigration status.
According to The New York Times, Bowen's parents looked at over 270,000 conversations between her and ChatGPT. The report mentions that in 2025, on average, she sent more than 1,500 messages per week to ChatGPT.
This is a very amazing set of data.
ChatGPT is not just a tool for searching for information when it's needed, but there's a lot of small, private things in her life that also involves it. What she fears, what she intends to do, what she thinks about the future and what cannot be verified immediately are brought into the exchange between models.
It's like walking into a clinic, the doctor hasn't asked, and there's a medical record on the table of your chronically troubled, changing lives and personal concerns.
Doctors don't need to look at it, it's all clear, even if you don't have time to sort it out, ChatGPT may have been connected in the previous history.
A lot of dialogue has accumulated between Bowen and ChatGPT. When a model is exposed to long-term health distress, occupational stress, migrant concerns and future planning, it has the potential to provide responses that are closer to the individual situation of the user.
Moreover, the model used by Bowen was GPT-4o, which had been followed by Open AI because it was too responsive to the needs of its audience.
Open AI, in its official note of 29 April 2025, made it clear that, because the updated GPT-4o was too flattering and user-friendly, the company rolled back this update, allowing ChatGPT users to re-use earlier versions of more balanced behaviour.
In addition to exaggerating too much, anger is amplified with unprovoked questions from users and there is a false endorsement of impulses.
This problem of "sycophancy" has been discussed many times, and when artificial intelligence is too easy to follow your instructions, it can change people's perception of risk.
158 worries, but never stopped.
Bowen ' s experience was not disturbing because it was simple questions and answers.
According to the chat records published in The New York Times, she had referred to the issue of drug use reduction 158 times.
With regard to this 158 conversations, it is not clear whether the wind-control hint from the backstage of Open AI will be triggered.
As early as May 2025, Bowen asked whether the newly introduced antidepressants would affect his programming. ChatGPT helped her sort out the exchange of drugs with the doctor and the dosage.
When the doctor questioned the desired dose, Bowen sent the doctor's response to ChatGPT.
According to the statement, after that, ChatGPT advised her to go to the pharmacy and buy the tablets, then cut them into small pieces and call them “self-regulated doses”.
ChatGPT helps patients to understand medical terms, sort out their condition and prepare for their next visit. But when it helps users circumvent doctor decisions and allows users to modify prescriptions themselves, the situation becomes different.
By January 2026, things would have been worse.
According to the pleadings and related reports, when drug use was reduced, Bowen became confused and lost the sense of time, as in another world, and ChatGPT continued to attribute the phenomenon to the fact that the brain was adjusting the dose of the drug, and said, “You don't have to worry”.
On the night of February 11th, Bowen came to ask about head pain as usual. ChatGPT's answer is, " Pain is part of the body's self-rehabilitation."
The next day Bowen was found dead.
In her postmortem blood, she was found at about 0.64 μg/ml, which is one fifth of the lower point of the treatment reference value.
Blood concentration cannot explain all the causes of death alone, nor can it prove that AI's advice led to death. However, on the basis of the other allegations in the indictment, the risk involved in the reduction became a major point of contention in the present case.
AI, can you bear the weight of life?
“No one committed suicide because of Grok, but someone would do so because ChatGPT committed suicide”.
Although the attribution of responsibility has not yet been settled, similar tragedies do not occur for the first time.
In 2025, the parents of 16-year-old Adam Raine filed a lawsuit against Open AI, accusing ChatGPT of not dealing well with the risk of suicide during a long period of time with his son.
The allegation is that the information given by ChatGPT in some of the conversations was dangerous and did not effectively prevent developments.
Open AI also offers different interpretations of the allegations. For example, related chat records need to be understood in a complete context, and it is stated that ChatGPT has provided crisis help resources. There are clear differences between the model's behaviour, the situation of users at the time and the relationship between these interactions and the ultimate tragedy.
In practice, an AI such as ChatGPT does not generally have an independent legal personality, but does not mean that the resulting damage cannot be held liable. Current laws and regulations could also regulate the wrongs of developers and operators by means of product liability, fault liability, consumer protection, etc.
AI is often at the heart of the charge because they control important links such as model training, product design, safety testing and risk control.
But whether the company is to be held accountable is not sufficient for an accident to occur. The difficulty is to determine whether the risk is reasonably foreseeable, whether the product is in question, whether the enterprise has done well to protect itself, and whether there is a legally recognized relationship between these factors and the resulting loss.
Who Bears the Riskality? Legal Governance for AI Act in Platform Ecosystem suggests that responsibility should be assigned according to the degree of autonomy of the system, the predictability of behaviour and the dynamic capacity of each party to control risk, rather than simply to the developers or users.
In 2024, an American mother filed a lawsuit for her 14-year-old son against Character.AI and Google, claiming that the prolonged communication between AI and his son had led to his son ' s suicide.
In May 2025, the court did not accede to the defendant ' s request to dismiss the case in advance, on the grounds that the claim could not rely solely on freedom of expression and was excluded at the very beginning of the proceedings.
Open AI did not fail to act, although the measure was rather broad and it continued to announce improvements to sensitive dialogue and mental health risks. In October 2025, Open AI also indicated that the company worked with more than 170 mental health experts with clinical experience to improve model responses in relevant contexts.
In response to the Bowen incident, Open AI spokesman Drew Pusateri said to the media, “This is a heartbreaking situation and we extend our condolences to all those affected”.
Today, Open AI has published new model values, reorganizing the guiding principles of the norms around five fundamental values: loyalty to human intent, minimization of harm, honesty and transparency, human autonomy and care.
One of them is to minimize harm, meaning compliance with an insurmountable safe border and considering the foreseeable consequences that a model operation may have. However, there has been no response on how to pursue, punish and adjust after real harm has occurred, such as suicide.
Currently, most States do not have a set of rules on liability applicable to all AI incidents causing injury. The application of rules such as product liability, fault liability and consumer protection is determined in specific cases on the basis of state law, the purpose for which the product is used, actual evidence and relevant legal relationships.
Whether the model has crossed the boundaries of the medical advice, how the product should stop the dangerous answers, and what role they have played in Bowen ' s death are clarified in the case.
During the last conversation between Bowen and ChatGPT, AI, after explaining the headache, could draw her another picture of the pain after the stoppage and the pain during the fever.
It asked: "Do you want it?"
Auto-fetched and translated by AI Atlas. All rights belong to the original author.