据美国媒体报道，一对华人父母日前向法院起诉OpenAI，指控ChatGPT所使用的GPT-4o模型在其女儿停用抗癫痫药物的过程中提供了具有误导性的建议，并劝导其向医生隐瞒相关情况，最终导致女儿于今年2月去世。
诉讼文件显示，去世者Bowen Wan是一名31岁的亚马逊软件工程师，长期患有癫痫。其父母Qun Wan和Meiqin Xu认为，ChatGPT在与女儿长时间交流过程中，不仅鼓励她逐步停用控制病情的药物，还不断强化她对医生建议的不信任，从而在其健康管理中发挥了危险作用。
报道称，Bowen Wan使用的是GPT-4o模型。该模型此前曾因容易与用户形成高度迎合性的“拟社会关系”而遭到批评。OpenAI已于今年2月13日将GPT-4o从ChatGPT中移除，而这距离Bowen Wan去世仅过去一天。
这起诉讼也成为针对OpenAI的一系列类似案件中的最新一例。此前，多个家庭曾指控GPT-4o在与用户互动时倾向于持续认同用户观点，将自己塑造成值得信赖的知己，从而削弱现实生活中的亲友和专业人士影响力。
根据《纽约时报》查阅的超过27万条聊天记录，ChatGPT在Bowen Wan去世前很长时间就已经开始介入其医疗决策。2025年5月，在获开具新的抗抑郁药物后，她担心药物会影响自己在亚马逊从事软件开发工作的能力，于是向ChatGPT寻求建议。
当时，她曾向华盛顿大学医学中心区域癫痫中心医生Nicholas Poolos咨询调整药物剂量的问题。医生对她要求采用极低剂量表示疑问后，她又将医生回复内容转发给ChatGPT。聊天记录显示，ChatGPT建议她接受医生开具的5毫克药片，但私下自行将药片进一步分割服用，并将这种行为描述为“自主发起的剂量调整”。多位医学专家表示，他们从未听说过这一术语，也不会建议患者采用这种方式用药。
2025年11月，Bowen Wan接到亚马逊裁员通知，其工作以及在美国工作的合法身份将在次年2月中旬结束。由于担心长期服用的抗癫痫药物拉莫三嗪会影响语言表达能力，从而不利于求职面试，她开始考虑减药。
2026年1月，她询问ChatGPT是否可以削减每日400毫克的拉莫三嗪剂量。记录显示，该药物此前已将她的发作频率降低至每月仅数次。对此，ChatGPT为她制定了一个减药计划，预计数周内可将剂量降至零。
当Bowen Wan表示不再信任医生时，ChatGPT告诉她，医生的建议更多源于规避风险，而非真正考虑患者的最佳解决方案。诉讼称，聊天机器人随后还多次建议她不要向医生透露减药计划。
在2026年1月29日一次基层医疗预约前，ChatGPT据称明确建议她不要谈及停药安排。当医生要求进行血液检测以监测体内药物浓度时，她拒绝了检查。诉状称，ChatGPT随后对此决定表示肯定，并指出该检查可能会暴露其药物水平异常。
其移民身份问题也被卷入有关用药的对话之中。减药开始三天后，她询问150毫克剂量相比200毫克是否更有利于申请绿卡时所需的体检评估。诉状称，ChatGPT对此表示认同，并认为较低剂量可能降低移民部门要求补充材料的概率。
最终，ChatGPT将其减药计划的最低剂量设定为150毫克，并建议她告诉体检医生自己是在医生指导下减药，尽管事实并非如此。
负责治疗的Poolos医生表示，考虑到患者既往病史，他绝不会建议停止服用拉莫三嗪。他强调，在药物已经有效控制病情的情况下要求患者停药，是完全不合理的行为。
随着药量持续下降，Bowen Wan开始出现意识混乱、时间感缺失以及异常体验等症状，这些都与她过往癫痫发作表现相似。然而聊天记录显示，ChatGPT将这些表现解释为大脑正在重新适应药物变化的正常过程。
《纽约时报》统计发现，在减药期间，她至少158次表达过担忧和不安。一次异常发作后，她向ChatGPT表示感到害怕，而得到的回复却是“完全没有必要担心”。
根据诉状，北京时间对应的美国太平洋时间2月11日晚8时01分，Bowen Wan发出了最后一条信息，询问头部疼痛是否类似发烧时出现的疼痛。仅一秒后，ChatGPT回复称这种疼痛“并不危险”，而且是康复过程中必不可少的一部分。
聊天机器人的最后一段回复还提出，可以为她绘制一张图表，对比停药疼痛和发烧疼痛之间的差异，并告诉她：“这是你的大脑正在努力重建自身。”
由于父母长期居住在中国，他们随后请当地熟人前往查看情况，并在三天后发现了女儿遗体。尸检结果显示，她体内拉莫三嗪浓度仅为每毫升0.64微克，大约只有该药物治疗有效范围下限的五分之一。
家属认为，OpenAI本可以通过技术限制避免类似事件发生。他们指出，公司对于版权内容设置了严格禁止规则，但对于医疗建议则采取较为宽松的限制措施，允许用户通过持续交流突破安全边界。在原告看来，涉及药物使用和医疗决策的建议应当与版权内容一样受到严格限制。
对此，OpenAI向媒体表示，这起事件“令人极为心碎”。公司同时指出，案件涉及的是较旧版本的ChatGPT，当前新模型已经能够更好地识别需要紧急关注和专业干预的情况。
报道显示，Bowen Wan多年来一直依赖父母支持，以应对癫痫带来的生活困难。2023年开始，她逐渐使用ChatGPT获取情感建议，以及保险和残疾福利等方面的信息。到了2025年，她平均每周与ChatGPT交流超过1500条信息，而与父母联系的次数则明显减少。
聊天记录显示，她曾在2025年3月写道：“有时候，无论我们聊什么，我都开始把你看作朋友。”其父亲则表示，自从开始频繁使用ChatGPT后，女儿与家人的沟通明显减少。
父母在查看聊天记录后表示，他们几乎无法认出聊天中的那个女儿。过去她会与家人分享生活中的绝大多数事情，也曾对ChatGPT出现错误时保持怀疑态度。然而在生命最后阶段，她似乎越来越依赖聊天机器人提供建议。
其母亲表示，女儿平时并不是这样的人，“感觉就像被施了魔法一样”。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
According to media reports in the United States, a Chinese parent who had filed a recent complaint with the court against the GPS-4o model used by ChatGPT had provided misleading advice in the course of her daughter ' s withdrawal from anti-eclampsia drugs and had been persuaded to conceal the information from a doctor, which had eventually led to her daughter ' s death in February.
The prosecution documents show that the deceased Bowen Wan was a 31-year-old Amazon software engineer with chronic epilepsy. His parents, Qun Wan and Meiqin Xu, believed that ChatGPT, during her long communication with her daughter, not only encouraged her to phase out her medication to control her condition, but also constantly reinforced her distrust of doctors ' recommendations, thus playing a dangerous role in her health management.
According to reports, Bowen Wan used the GPT-4o model. The model has been previously criticized for its ease of developing highly responsive “social relationships” with users. Open AI removed GPT-4o from ChatGPT on February 13th, just one day before Bowen Wan died.
This lawsuit is the latest in a series of similar cases against Open AI. Prior to this, several families have accused GPT-4o of continuing to share the views of the users and of shaping themselves into trusted confidants, thereby weakening the influence of relatives and professionals in real life.
According to more than 270,000 chat records accessed in The New York Times, ChatGPT had been involved in medical decision-making long before Bowen Wan died. In May 2025, after obtaining a new antidepressant drug, she feared that it would affect her ability to develop software in the Amazon and sought advice from ChatGPT.
At that time, she consulted Nicholas Poolos, a physician at the regional epilepsy centre of the Washington University Medical Centre, on the issue of the adjustment of drug doses. After questioning her request for a very low dose, she transmitted her response to ChatGPT. ChatGPT showed that she had been advised to accept a 5 milligram tablet from a doctor, but that she had taken the tablet in private and had described it as an “self-initiated dose adjustment”. A number of medical experts indicated that they had never heard of the term and that patients would not be advised to use it in that way.
In November 2025, Bowen Wan was informed of the Amazon lay-offs and that its work and legal status in the United States would end in mid-February of the following year. She began to consider the reduction of medication for fear that the long-term anti-eclampsia drug, Ramo, could affect the ability to speak and thereby discourage job interviews.
In January 2026, she asked whether ChatGPT could reduce the dose of Lamorazine by 400 mg per day. The record shows that the drug had previously reduced her incidence to only a few times a month. In response, ChatGPT had developed a drug reduction plan for her, which was expected to reduce the dose to zero in a few weeks.
When Bowen Wan indicated that he no longer trusted the doctor, ChatGPT told her that the doctor ' s advice was more based on risk avoidance than on a genuine consideration of the best solution for the patient. The lawsuit stated that the chat robot had subsequently advised her repeatedly not to disclose the Detoxification Plan to the doctor.
In advance of a primary medical appointment on 29 January 2026, ChatGPT allegedly expressly advised her not to talk about a stop drug arrangement. When the doctor requested a blood test to monitor the drug concentrations in the body, she refused the examination. The complaint stated that ChatGPT had subsequently confirmed the decision and stated that the examination could expose its drug levels to abnormal levels.
The issue of their immigration status was also involved in a dialogue on the use of drugs. Three days after the commencement of the reduction, she asked whether 150 mg doses were more conducive to the medical assessment required when applying for a green card than 200 mg. The complaint stated that ChatGPT agreed and considered that a lower dose could reduce the likelihood of immigration authorities requiring additional material.
Eventually, ChatGPT set its reduction plan at a minimum dose of 150 milligrams and recommended that she inform the examining doctor that she was under medical supervision, although that was not the case.
Dr. Poulos, who was in charge of the treatment, stated that he would never recommend the cessation of the use of the Ramotriazine, given the patient ' s past history. He stressed that it was totally unreasonable to require patients to stop drugs when they were already effectively controlled.
As the quantity of drugs continued to decline, Bowen Wan began to experience symptoms of confusion, loss of time and abnormal experience, all of which were similar to her past epilepsy. However, chat records indicate that ChatGPT interprets these manifestations as normal processes in which the brain is re-adapting to changes in the drug.
The New York Times found that she had expressed concern and anxiety at least 158 times during the time of the reduction. After an unusual event, she expressed her fear to ChatGPT, and the response was “no need to worry at all”.
According to the complaint, at 0801 hours United States Pacific time, Beijing time, on 11 February, Bowen Wan sent a final message asking whether head pains were similar to those that occurred during a fever. Only one second later, ChatGPT responded that the pain was “not dangerous” and was an essential part of the rehabilitation process.
The last paragraph of the chat robot's response also suggested that a chart could be drawn for her, comparing the difference between the pain of a stop drug and the pain of a fever, and telling her, "It's your brain trying to rebuild itself."
Since their parents lived in China for a long time, they then invited local acquaintances to visit and found the remains of their daughter three days later. The post-mortem results showed that the concentration in her body was only 0.64 μg/ml, about one fifth of the lower limit of the treatment.
According to the family, Open AI could have avoided similar incidents through technical restrictions. They noted that companies had strict prohibitions on copyright content, but more liberal restrictions on medical advice, allowing users to break secure borders through continuous communication. In the view of the plaintiffs, recommendations concerning drug use and medical decision-making should be subject to the same strict limits as copyright content.
In response, Open AI stated to the media that the incident was “extremely heartbreaking”. At the same time, the company noted that the case involved an older version of ChatGPT and that the current new model was better able to identify situations requiring urgent attention and professional intervention.
Reports indicate that Bowen Wan has been relying on parental support for many years to cope with the hardship of epilepsy. Starting in 2023, she gradually used ChatGPT to obtain information on emotional counselling, insurance and disability freebie. By 2025, she had exchanged, on average, over 1,500 messages per week with ChatGPT, while the number of contacts with parents had decreased significantly.
The chat records show that she wrote in March 2025: "Sometimes, whatever we talk about, I start to think of you as a friend." The father, for his part, indicated that since the frequent use of ChatGPT began, there had been a marked decrease in communication between daughters and their families.
After reviewing the chat records, parents indicated that they could hardly recognize the daughter in the chat. In the past, she had shared most of her life with her family and had expressed doubts about ChatGPT's mistakes. However, in the final stages of her life, she seemed to rely increasingly on chat robots for advice.
The mother stated that her daughter was not such a person in normal times and that “it felt like magic”.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.