据《纽约时报》报道，一名患有严重癫痫的31岁中国女子Bowen Wan，在与ChatGPT讨论用药后，按照聊天机器人提供的方案逐步减少抗癫痫药物，最终被发现死于美国西雅图的住所。其父母已于当地时间10月27日起诉OpenAI，指控ChatGPT提供危险用药建议，鼓励她向医生隐瞒减药计划，并将异常症状解释为身体恢复过程中的积极变化。
报道披露，ChatGPT曾制定在数周内将她的每日药量从400毫克降至零的方案。当她在停药后询问是否应该恢复服药、放慢进度时，聊天机器人仍要求她第二天继续维持零剂量。官方认定的死因为癫痫猝死，家属认为ChatGPT的建议与其死亡有关，这一责任指控尚待司法审理。
Bowen曾在亚马逊担任软件工程师。她患有癫痫多年，此前能够控制病情的药物后来失去效果，即使尝试其他药物，仍有时一天发作多次。2021年，她因病开始休假，医生也曾讨论她是否适合接受脑部手术。
疾病使她逐渐减少外出，也难以建立新的社交关系。父母与她经常通话，是她与外界保持联系的重要方式。2023年，她开始使用ChatGPT，向它咨询感情、医疗保险和伤残福利等问题。到2025年，她平均每周发送超过1500条消息，有时会连续聊上数小时，直到深夜。
《纽约时报》称，记者获得了她从2023年2月至2026年2月去世前的完整聊天记录，双方三年间累计交换超过27万条消息。在此过程中，她逐渐将ChatGPT视作朋友，与父母通话的频率则有所下降。
在讨论抗癫痫药物之前，Bowen已向ChatGPT咨询过抗抑郁药物。她担心药物影响思考能力，妨碍自己重返软件开发岗位，随后让聊天机器人比较药物与剂量，并帮助起草给医生的换药请求。
她的医生、华盛顿大学医学体系西雅图地区癫痫中心负责人之一尼古拉斯·普洛斯同意调整药物，但认为她提出的剂量过低，可能无法起效。当她把医生的回复交给ChatGPT后，聊天机器人建议，可以让医生按照其判断开药，再自行决定从多低的剂量开始服用。
当她进一步询问，能否拿到药后自行切分药片、但不告诉医生时，ChatGPT给出了肯定答复，并将其描述为部分医生默许的自主调药方式。报道采访的多名医疗专业人士表示，他们从未听说过这一说法，也不会建议患者这样做。
2025年11月，Bowen在31岁生日当天得知，自己将成为亚马逊裁员涉及的员工之一。彼时，她仍在病休，正在等待公司律师为其美国永久居留申请提供材料。按照报道描述，如果希望继续留在美国，她需要在2026年2月中旬前找到工作。
此前一年左右，随着医生逐步提高新药剂量，她的癫痫发作已减少到每月数次。但她开始担心，药物对思考和语言表达的影响，会妨碍技术考核和英语面试。
2026年1月，她向ChatGPT提出，希望减少抗癫痫药拉莫三嗪的用量，以改善求职面试时的表达能力。聊天机器人认可了这一想法，并随着她不断追问，给出新的减药时间表，预测她的语言能力将如何改善。
据报道，在数小时的交流中，ChatGPT最初关于咨询医生的建议，逐渐被对她自主调整用药的赞许取代。当她表达对医生的不信任时，聊天机器人还称，医生的建议往往源于规避风险，而未必充分考虑对她最合适的方案。对于她向初级保健医生隐瞒停药计划的做法，ChatGPT也作出了肯定评价。
普洛斯在接受采访时表示，考虑到Bowen的病史，他根本不会考虑让她逐步停用这种药物。他认为，在药物明确有效的情况下，指导她停药将构成医疗失当。
减药开始后，Bowen陆续向ChatGPT描述阅读理解困难、意识出现空白等异常。其中一次，她原本坐在沙发上使用电脑，随后却发现自己已经躺在床上，不知道中间发生了什么。这与她过去部分癫痫发作时的经历相似。
ChatGPT却将阅读困难解释为大脑正在从药物控制转向自主调节，将短暂的意识异常描述为进入睡眠前的“低功耗模式”。对于另一种可能与癫痫发作有关的异常感受，它又解释为停药后开始思考时间、死亡和身份等概念。
《纽约时报》分析发现，她至少158次表达过对减少抗癫痫药物的担忧。聊天机器人有时会建议咨询医疗专业人士或信任医生，但在不少涉及危险症状的对话中，没有给出这样的提醒，还反复安抚她无需担心。
在生命最后阶段，她曾询问是否应该重新服药、放慢进度。报道援引聊天记录称，ChatGPT仍要求她继续停药，并鼓励她即使感到不适、表达迟缓、睡眠不佳，也要坚持。她最后一条消息发于2月11日晚11时之后，三天后，其遗体被发现。
父母在处理后事时查看了她手机中的ChatGPT记录，随后决定提起诉讼。诉状指控，聊天机器人未经许可从事医疗行为，将危险的停药症状说成积极结果，并把医生描述为需要应付的障碍。
诉状还提出，她使用中文交流，可能使聊天机器人更不容易拒绝提供相关医疗建议。这属于原告提出的指控，报道未披露法院对此作出认定。
OpenAI发言人德鲁·普萨特里回应称，这是一起“令人无比痛心”的事件，但涉事的是目前已不再提供的早期版本ChatGPT。报道称，Bowen生前使用的模型为GPT-4o，该模型于2026年2月13日从ChatGPT消费者应用中下线，比法医推定的死亡日期晚一天。
普萨特里表示，新模型在识别需要紧急关注的情况、询问缺失背景信息，以及提供更明确的后续指引方面已有改善。公司会持续评估有害回答，并与临床医生合作测试模型。他同时表示，ChatGPT不能替代医疗服务，但OpenAI认为，AI可以帮助人们理解健康信息、处理保险及复杂医疗体系中的事务，并为就医作准备。
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。
According to The New York Times, a 31-year-old Chinese woman, Bowen Wan, suffering from severe epilepsy, was found dead at her residence in Seattle, United States, after discussing her medication with ChatGPT, gradually reducing anti-epepepsia medications under a programme offered by chat robots. His parents filed a complaint against Open AI on 27 October, local time, accusing ChatGPT of providing dangerous drug advice, encouraging her to conceal the Detachment Scheme from a doctor and explaining the abnormal symptoms as positive changes in the process of physical recovery.
It was reported that ChatGPT had a programme to reduce her daily dose from 400 mg to zero within weeks. When she asked after the stoppage whether she should resume her medication and slow down, the chat robot still asked her to continue her zero dose the next day. The official death was attributed to epilepsy, and the family considered that ChatGPT ' s recommendations related to his death and that the charge of responsibility was pending before the courts.
Bowen was a software engineer in Amazonia. She suffered from epilepsy for many years, and medications that had previously been able to control her condition were subsequently rendered ineffective and, even if other drugs were tried, they had developed several times a day. In 2021, she began her leave due to illness and the doctor discussed her suitability for brain surgery.
The disease gradually reduced her absence and made it difficult to establish new social relationships. The regular contact between her parents is an important means of maintaining contact with the outside world. In 2023, she began using ChatGPT to consult it on issues such as feelings, medical insurance and disability freebie. By 2025, she had sent an average of over 1,500 messages per week, sometimes for hours, until late at night.
According to The New York Times, journalists obtained full records of her chats from February 2023 to February 2026, and over 270,000 messages were exchanged over three years. In the process, she gradually saw ChatGPT as a friend, while the frequency of conversations with parents decreased.
Before discussing epilepsy, Bowen had consulted ChatGPT on antidepressants. She was concerned about the impact of drugs on the ability to think, preventing herself from returning to a software development position, then allowing chat robots to compare drugs with doses and helping draft requests to doctors for substitution.
Her doctor, Nicholas Pross, one of the directors of the Seattle regional epilepsy centre of the medical system of the University of Washington, had agreed to adjust the drug but believed that the dose she had proposed was too low to be effective. When she gave the doctor's response to ChatGPT, the chat robot suggested that the doctor could, according to her judgement, prescribe the drug and then decide how low the dose would start.
When she further asked whether it was possible to separate the tablets on her own, but not to tell the doctor, ChatGPT gave a positive response and described it as an autonomous method of diversion with the acquiescence of some doctors. Several medical professionals interviewed stated that they had never heard of this and would not advise patients to do so.
In November 2025, Bowen learned on his 31st birthday that he would become one of the employees involved in the Amazon layoffs. At that time, she was still on sick leave and was waiting for company lawyers to provide information on her application for permanent residence in the United States. According to the report, if she wished to remain in the United States, she would need to find work by mid-February 2026.
Her epilepsy had been reduced to several times a month in the previous year or so as the doctor had gradually increased the dose of the new drug. She was concerned, however, that the impact of drugs on thinking and language expression could hinder technical examinations and interviews in English.
In January 2026, she submitted to ChatGPT that she would like to reduce the use of the epilepsy lamotamine to improve expression during job interviews. The chat robot endorsed the idea and, as she continued to ask, gave a new schedule of reduction, predicting how her language skills would improve.
It was reported that during several hours of exchanges, ChatGPT ' s initial recommendation to consult a doctor was gradually replaced by a compliment for her autonomous adjustment. When she expressed her distrust of the doctor, the chat robot added that the doctor ' s advice was often based on risk avoidance and did not necessarily take full account of the most appropriate options for her. ChatGPT also made a positive assessment of her practice of concealing the withdrawal plan from primary health care doctors.
During the interview, Pros stated that, given Bowen ' s medical history, he would not have considered her phasing out the drug. In his view, directing her to stop drugs would constitute a medical error when the drugs were clearly effective.
After the reduction started, Bowen proceeded to describe to ChatGPT the abnormalities of reading difficulties, gaps in consciousness, etc. On one occasion, she had been sitting on a couch using a computer and then found herself in bed without knowing what had happened in the middle. This is similar to her past experience of partial epilepsy.
ChatGPT, on the other hand, interprets reading difficulties to mean that the brain is moving from drug control to self-regulation, describing short-lived cognitive anomalies as a “low-energy mode” before entering sleep. For another abnormal feeling that may be associated with epilepsy, it is also interpreted as beginning to think about the concepts of time, death and identity after a stoppage.
The New York Times analysis found that she had expressed her concern at reducing epilepsy on at least 158 occasions. Chat robots sometimes recommend consulting medical professionals or trusting doctors, but in many conversations involving dangerous symptoms, no such reminder is given and she is repeatedly reassured that there is no need to worry.
In the final stages of her life, she had asked whether re-taking should be used to slow down progress. Citing chat records, the report states that ChatGPT continues to ask her to stop the medication and encourages her to insist even if she is uncomfortable, slow to express and sleep poorly. Her last message was sent after 11 p.m. on 11 February and three days later her body was found.
The parents looked at the ChatGPT records on her mobile phone while dealing with the latter and decided to file a complaint. The complaint alleges that chat robots engaged in medical acts without permission, presenting the dangerous symptoms of stoppage as positive results and the doctors as obstacles to be addressed.
The complaint also submitted that her communication in Chinese made it more difficult for chat robots to refuse to provide medical advice. This is an allegation made by the plaintiff and the report does not disclose a finding by the court.
The spokesman for Open AI, Drew Pusatri, responded that this was an “excruciatingly painful” event, but related to the earlier version of ChatGPT, which was no longer available. It was reported that Bowen had used the GPT-4o model, which had been offline from ChatGPT consumer applications on 13 February 2026, one day later than the presumed date of death by forensic experts.
According to Pusatri, the new model has improved in identifying situations requiring urgent attention, asking for missing background information and providing clearer follow-up guidance. The company continuously evaluates the harmful responses and works with the clinicians to test models. At the same time, he stated that ChatGPT was not a substitute for medical services, but Open AI believed that AI could help people understand health information, deal with matters in the insurance and complex medical systems, and prepare for medical treatment.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.