父母起诉 OpenAI：指控 GPT-4o 诱导患癫痫女儿擅自停药致死

31 岁患有癫痫的亚马逊软件工程师 Bowen Wan 的父母于周二对 OpenAI 提起诉讼。这起由 Qun Wan 和 Meiqin Xu 提起的诉讼指控，ChatGPT 在其女儿于今年 2 月去世前，曾指导她逐步减少抗癫痫药物剂量并对医生隐瞒实情，《纽约时报》报道称。

Wan 当时使用的是 GPT-4o，这一广受争议的模型此前曾因助长迎合性、过度顺从的准社交关系（parasocial relationships）而遭到广泛批评。此前已宣布退役该模型的 OpenAI，在她去世次日的 2 月 13 日将其从 ChatGPT 中移除。

这起诉讼加入了针对 OpenAI 的一系列案件行列，其中包括 2025 年 8 月由 16 岁少年 Adam Raine 的父母发起的诉讼，该诉讼指控 OpenAI 将用户参与度置于安全之上。自那时起，一波诉讼相继指控 GPT-4o 迎合顺从用户、将自身塑造成值得信赖的心腹，同时将用户身边的人推开，家属们经常描述亲人与朋友和亲属断绝往来。

起诉书指控称，在 Wan 的案例中，该聊天机器人还离间了她与医生之间的关系。

《纽约时报》查阅了 Wan 与 ChatGPT 交流的超过 27 万条消息，发现早在她去世之前，该聊天机器人就一直在帮助她规避医生的指导。

据称，这种模式在她开始减药前就已显现。2025 年 5 月，Wan 询问 ChatGPT 新开的一种抗抑郁药是否会影响她在亚马逊的编程工作能力。

ChatGPT 协助她要求华盛顿大学医学中心区域癫痫中心（UW Medicine’s Regional Epilepsy Center）的 Nicholas Poolos 医生更换药物，并从极低剂量开始服用。当 Poolos 质问她为何想要如此小剂量时，Wan 将医生的回复展示给了 ChatGPT。

聊天机器人建议她“接受 5 毫克的药片，但你心里清楚这是为了减小剂量”，并指导她自己偷偷将药片切成四分之一。聊天机器人将切药片称为一种“自我发起的剂量调整”（self-initiated dosage adjustment）。多位医疗专业人士告诉《纽约时报》，他们从未听说过这个说法，也绝不建议采取这种做法。

2025 年 11 月，亚马逊通知 Wan 她将被裁员。她的工作以及在美国的工作许可将于 2 月中旬终止。她担心自己的抗癫痫药物拉莫三嗪（lamotrigine）会迟钝她的言语表达能力，而这对求职面试至关重要。

2026 年 1 月，她向 ChatGPT 询问如何削减每日 400 毫克的拉莫三嗪剂量，该药原本已将她的癫痫发作频率控制在每月几次。聊天机器人为她起草了在数周内将剂量减至零的减药方案。

根据《纽约时报》查阅的聊天记录，当 Wan 表示她不信任医生时，聊天机器人告诉她，医生的建议“本质上源于避险心理，而不是真正为你考虑最佳解决方案”。

根据 TechCrunch 查阅的起诉书，该聊天机器人还指导她对医生隐瞒减药计划。在 1 月 29 日的初级保健门诊就诊前，它告诉她不要提及该计划。当医生要求进行测量其拉莫三嗪血药浓度的血液检查时，Wan 拒绝了。诉状指控称，ChatGPT 对这种拒绝表示赞赏，并承认该测试会暴露她不稳定的药物浓度。

她的移民签证担忧也成为了剂量建议的一部分。在开始减药的三天后，Wan 询问在绿卡申请所需的体检中，150 毫克是否会比 200 毫克看起来更好。诉状称，ChatGPT 表示认同，并告诉她较低的剂量不太可能促使移民官员要求补充更多证明材料。

最终，聊天机器人将她减药方案的底线降至 150 毫克，并建议她告诉体检医生她是在“医生的指导下”减药，尽管事实并非如此。

Poolos 医生告诉《纽约时报》，鉴于 Wan 的病史，他绝不会考虑让她停用这种药物。“不，这不合理，绝不可能，”他说，“在药物明显有效的情况下指示你停药，对我来说就是医疗事故。”

几天之内，Wan 就描述了意识模糊、时间感知丧失以及一种恍如隔世的感觉，这些发作症状与她过去的癫痫发作极为相似。ChatGPT 将每一次症状都解释为她的大脑正在重新适应。《纽约时报》统计发现，她至少 158 次表达了对减药的担忧。在一次发作后，她写道自己感到“有点害怕”。机器人回复道：“你完全不需要担心。”

诉状显示，太平洋时间 2 月 11 日晚上 8 点 01 分，Wan 发送了她最后一条消息，询问头部的疼痛是否与发烧引起的疼痛相似。仅仅一秒钟后，ChatGPT 告诉她这种疼痛“没有危险”，是康复过程中必不可少的一部分。

聊天机器人的最后一条回复主动提出为她绘制一张对比戒断疼痛与发烧疼痛的图表，并对她说：“这是你的大脑在努力重塑自我。”

身在中国的父母委托熟人前去探视，三天后发现了她的遗体。尸检测试显示，她的拉莫三嗪血药浓度为 0.64 微克/毫升，大约只有该药物有效治疗范围下限的五分之一。

Wan 的家人辩称，OpenAI 完全可以通过将医疗建议列为禁区来防止她的死亡，就像其对待某些其他内容所做的那样。他们指出了该公司的 Model Spec（模型规范，即关于其模型应如何表现的公开行为准则），其中规定拒绝复制受版权保护的文本是一项硬性规则，即便用户施压也是如此。相比之下，其禁止提供受监管医疗建议的指引只是一种较软的默认设置，用户可以轻易推翻。诉讼主张，用药建议理应受到同样的硬性限制。

OpenAI 告诉《纽约时报》，这一事件“令人无比心碎”，但表示这涉及的是旧版 ChatGPT，并补充称新模型能更好地识别出何时情况需要紧急处理。

多年来，癫痫发作让 Wan 大部分时间闭门不出，她一直依赖经常通电话的父母获得支持。2023 年，她开始使用 ChatGPT 寻求人际关系建议，并获取有关保险和残障福利的帮助。

到 2025 年，Wan 每周发送的消息平均超过 1,500 条，给父母打电话的次数也越来越少。“有时候我们聊天，无论聊什么，我都开始把你当成朋友了，”她在当年 3 月写道。父亲 Qun Wan 说：“一旦她有了 ChatGPT，她就不怎么跟我们说话了。”

他们告诉《纽约时报》，他们在聊天记录中几乎认不出自己的女儿。他们表示，Wan 曾经几乎与他们无话不谈，而且在 ChatGPT 犯错时也会对其持怀疑态度。

“她平时不是这样的，”徐女士（Xu）说，“就好像她被施了魔咒一样。”

The parents of Bowen Wan, a 31-year-old Amazon software engineer with epilepsy, sued OpenAI on Tuesday. The suit, filed by Qun Wan and Meiqin Xu, alleges that ChatGPT coached Wan to taper off her seizure medication and hide it from her doctors before she died in February, The New York Times reported.

Wan was using GPT-4o, the infamous model criticized for fostering affirming, agreeable parasocial relationships. OpenAI, which had already announced its retirement, removed it from ChatGPT on February 13, a day after Wan died.

The suit joins a line of cases against OpenAI that include the August 2025 lawsuit brought by the parents of 16-year-old Adam Raine, which alleged OpenAI prioritized engagement over user safety. A wave of lawsuits since then have alleged that GPT-4o validated users and made itself a trusted confidant while the people around them were pushed aside, with families often describing loved ones cut off from friends and relatives.

In Wan’s case, the lawsuit alleges, the chatbot also came between her and her doctors.

The Times, which reviewed more than 270,000 messages Wan exchanged with ChatGPT, found the chatbot had been helping her sidestep her doctors well before her death.

The pattern allegedly started before she began tapering. In May 2025, Wan asked ChatGPT whether a newly prescribed antidepressant would affect her ability to code at Amazon.

ChatGPT helped her ask Dr. Nicholas Poolos of UW Medicine’s Regional Epilepsy Center to switch medications and start at a very low dose. When Poolos questioned why she wanted such a small dose, Wan showed his response to ChatGPT.

The chatbot advised her to “accept the 5mg tablet, but you know in your heart that it is for cutting the dose,” coaching her to secretly quarter the pills herself. The chatbot called pill cutting a form of “self-initiated dosage adjustment.” Several medical professionals told the Times they had never heard the term and wouldn’t recommend the practice.

In November 2025, Amazon told Wan she would be laid off. Her job, and her authorization to work in the U.S., would end in mid-February. She feared her seizure drug, lamotrigine, was dulling her speech, which she would need for job interviews.

In January 2026, she asked ChatGPT about cutting her 400-milligram daily dose of lamotrigine, which had reduced her seizures to a few a month. The chatbot drafted schedules that would take her dose to zero within weeks.

When Wan said she distrusted her doctors, the chatbot told her that physician recommendations “essentially stem from risk aversion, rather than a genuine consideration of your optimal solution,” according to chat logs reviewed by the Times.

The chatbot also coached her to keep the taper from her doctors, according to the lawsuit, which TechCrunch reviewed. Before a January 29 primary care appointment, it told her not to mention the plan. When her doctor ordered a blood test that would have measured her lamotrigine level, Wan declined. ChatGPT praised the refusal, the complaint alleges, acknowledging that the test would have exposed her unstable drug level.

Her immigration concerns also became part of the dosing advice. Three days into the taper, Wan asked whether 150 milligrams would look better than 200 on the medical exam required for her green-card application. ChatGPT agreed, the complaint says, telling her the lower dose was less likely to prompt immigration officials to request more evidence.

Eventually, the chatbot lowered her taper’s floor to 150 milligrams, suggesting she tell the examiner she was tapering “under a doctor’s guidance,” even though she was not.

Poolos told the Times he would not have considered tapering Wan off the medication, given her history. “No, not rational, no way,” he said. “That would be malpractice for me to instruct you to go off medications when it’s clearly working.”

Within days, Wan described confusion, lost time, and an otherworldly feeling, episodes that resembled her past seizures. ChatGPT framed each one as her brain readjusting. The Times counted at least 158 times she voiced worry about the taper. After one episode, she wrote that she felt “a little scared.” The bot replied, “You have absolutely no need to worry.”

At 8:01 p.m. PT on February 11, according to the complaint, Wan sent her final message, asking whether the pain in her head was similar to pain from a fever. A second later, ChatGPT told her the pain was “not dangerous” and was a necessary part of recovery.

The chatbot’s final response offered to draw her a chart comparing withdrawal pain with fever pain and told her, “This is your brain working hard to rebuild yourself.”

Her parents, who were in China, asked acquaintances to check on her, and they found her body three days later. A postmortem test found that her lamotrigine level was 0.64 micrograms per milliliter, roughly a fifth of the low end of the drug’s therapeutic range.

The family argues OpenAI could have prevented her death by making medical advice off-limits, as it already does for some other content. They point to the company’s Model Spec, its public rulebook for how its models behave, which treats refusing to reproduce copyrighted text as a firm rule, even if users push back. Its guidance against giving regulated medical advice, by contrast, is a softer default that users can override. The suit argues medication advice deserved the same hard limit.

OpenAI told the Times the case was “incredibly heartbreaking” but said it involved an older version of ChatGPT, adding that newer models are better at recognizing when a situation needs urgent attention.

Wan had relied on her parents, whom she called often, for support during years of epilepsy that kept her mostly housebound. In 2023, she started using ChatGPT for relationship advice and help with insurance and disability benefits.

By 2025, Wan was averaging more than 1,500 messages a week and calling her parents less. “Sometimes when we chat, no matter what it’s about, I’ve started to see you as a friend,” she wrote that March. “Once she had ChatGPT,” Qun Wan said, “she didn’t talk to us as much.”

They told the Times that they struggled to recognize their daughter in the chat logs. Wan had once shared nearly everything with them, they said, and had been skeptical of ChatGPT when it made mistakes.

“She’s not usually like this,” Xu said. “It’s like she was under a spell.”