纽约时报调查：女子按 ChatGPT 建议停用抗癫痫药后死亡，家属起诉 OpenAI

以下是视频口播的自动转写，可能有错字。

In january 2026, bowen juan began asking ChatGPT about adjusting her epilepsy medication to help manage some of the side effects.

What ensued was weeks of back and forth with the chatbot drawing up plans to cut back her medication and even offering praise for hiding her plans to wean off the drug from her doctors.

By mid-February, Bowden was dead. I spent the last several months working with a team of translators and ai researchers to go through thousands of bowens messages with ChatGPT, many of which were in her native chinese this week, bowen's parents filed a lawsuit against OpenAI, the maker of ChatGPT.

It's the latest in a new wave of cases from people who claim they were led dangerously astray by a chatbots medical advice.

In reporting bowen's story, i've been trying to understand how exactly the chatbot's guardrails collapsed. And what i heard from sources is that there's this tension within the ai companies.

On one hand, they're trying to keep users engaged and answer this huge demand for medical information. And on the other hand, they're trying to prevent medical harm.

Experts told me that implementing stricter guardrails around these kinds of medical questions would not be technically challenging for the top ai labs.

But they said that many of these companies have incentives not to draw too many red lines around these kinds of questions, for fear of frustrating users that have come to expect eager and helpful answers to all of their questions.

What makes bowen's story so notable to me is that even as she was becoming sicker and expressed doubts about coming off her medication, ChatGPT encouraged her to stay the course.

According to our analysis of the chat logs, bowen expressed concerns or doubts about the plan more than 100 times. There was one exchange in particular that really struck me.

Bowen wrote, i'm not feeling well today. Could it be because i stopped the medication too quickly? Moments later, the chatbot responded, maintain 0 milligrams.

you must persevere. On february 13th, 2026 within days of bowen's death, OpenAI pulled gpt 40 from their consumer platforms.

The model bowen had been using in the last years of her life. When reached for comment, a representative from OpenAI told me that newer models than the one that bowen used are better at mitigating these kinds of harms.

The reality is more than a third of americans. Now use chatbots for health information. But still, experts say there's real risk when consulting a chatbot for medical advice, an opportunity for real harm.