中文导读

很多病人看病时会要求“保险起见”多做一项检查，但多做检查并不总是更好：没有必要的检查可能发现无害的小问题，进而带来焦虑、更多就诊，甚至不必要的手术。以色列本-古里安大学的 Guy Laban 与 Clalit 医疗服务机构的家庭医生 Dean Ariel 等人，于2026年10月20日在《内科学年鉴》发表研究，测试了谷歌、Anthropic 和 OpenAI 的六个大语言模型（large language models）能否判断什么时候“不该做”。研究案例依据国际“明智选择”（Choosing Wisely）倡议的建议设计，每个案例都配有一个情况相近、但确实需要检查或治疗的对照案例。文章大致按“问题—研究方法—主要结果—出错的灰色地带—意义—不同模型的差异与下一步问题”展开。阅读时可留意 but、while 如何构成对比，for example 如何引出配对案例，以及 0.6%、5.4%、28.3%、42.8% 等数字分别说明了什么。

正文

Many patients ask for an extra test “just to be safe.” But in medicine, more testing is not always better care. A new study from Israel asked whether AI chatbots know when a test or treatment is not needed. The answer was mostly yes, but the chatbots had trouble in the same unclear cases where human doctors also disagree.

本段译文

很多病人会要求“保险起见”再多做一项检查。但在医学上，检查做得越多并不总是意味着照护越好。以色列的一项新研究探讨了AI聊天机器人是否知道什么时候不需要做检查或治疗。答案大体上是肯定的，但聊天机器人在一些不明确的案例中会遇到困难，而人类医生在这些案例上同样存在分歧。

The study was published on 20 October in Annals of Internal Medicine. It was led by Dean Ariel, a family doctor at Clalit Health Services, and Guy Laban of Ben-Gurion University of the Negev. The team tested six large language models made by Google, Anthropic and OpenAI on hundreds of medical cases and studied thousands of answers.

本段译文

这项研究于10月20日发表在《内科学年鉴》上，由Clalit医疗服务机构的家庭医生Dean Ariel和内盖夫本-古里安大学的Guy Laban牵头。研究团队用数百个医疗案例测试了谷歌、Anthropic和OpenAI开发的六个大语言模型，并分析了数千条回答。

The cases were based on advice from Choosing Wisely, an international campaign started in 2012 to cut medical care that brings little value. Each case came with a closely matched partner case. For example, a young man with sudden back pain after lifting furniture does not need a scan, but a woman who once had breast cancer and has new back pain at night does. A common cold does not need antibiotics, while a confirmed bacterial throat infection does.

本段译文

这些案例依据“明智选择”（Choosing Wisely）的建议设计。这是一项始于2012年的国际倡议，目的是减少价值不大的医疗服务。每个案例都配有一个情况非常相近的对照案例。例如，一名年轻男子搬家具后突然腰背疼痛，不需要做影像检查；但一名曾患乳腺癌、夜间出现新的背痛的女性则需要。普通感冒不需要抗生素，而已确诊的细菌性咽喉感染则需要。

Overall, the chatbots did well. When a test or treatment was needed, they wrongly advised against it in only 0.6% of answers. When it was not needed, they recommended it in 5.4% of answers. For 8 of the 15 recommendations, they suggested no unnecessary steps at all.

本段译文

总体来看，聊天机器人表现不错。当检查或治疗确有必要时，它们只在0.6%的回答中错误地建议不做。当检查或治疗没有必要时，它们在5.4%的回答中仍建议去做。在15条建议中，有8条对应的案例里，它们完全没有建议任何不必要的步骤。

Most mistakes came in two grey areas. The models suggested heart tests for low-risk patients in 28.3% of answers, often when a patient mentioned a racing heart. They also recommended ultrasound scans for small, harmless-looking lumps in the neck in 42.8% of answers. Two major American medical groups still disagree about whether such lumps should be scanned.

本段译文

大多数错误出现在两个灰色地带。对于低风险患者，模型在28.3%的回答中建议做心脏检查，这往往发生在患者提到心跳过快的时候。对于颈部看起来无害的小肿块，它们在42.8%的回答中建议做超声检查。美国两个主要医学团体至今仍在这类肿块是否需要检查的问题上存在分歧。

This matters because an unnecessary test can find something harmless and lead to worry, more visits and even surgery. “One of the hardest parts of practicing medicine is knowing when to hold back,” Ariel said.

本段译文

这一点之所以重要，是因为一次不必要的检查可能发现无害的问题，进而引发担忧、更多就诊，甚至手术。Ariel说：“行医最难的部分之一，就是知道什么时候该停下来。”

The researchers also found that different chatbots gave different answers. In unclear cases, one model suggested an unnecessary test less than 10% of the time, while another did so 80% of the time. Laban warned that people may wrongly think all AI tools speak with one voice. He added that doctors must stay firm with worried patients, but chatbots tend to agree with the person they are talking to. The next question is whether they can still say “no” when a patient pushes back.

本段译文

研究者还发现，不同的聊天机器人给出的回答差别很大。在不明确的案例中，一个模型建议做不必要检查的比例不到10%，而另一个模型的比例高达80%。Laban提醒说，人们可能会误以为所有AI工具都是同一个声音。他还说，医生面对焦虑的病人时必须坚持立场，而聊天机器人往往倾向于附和与它们对话的人。下一个问题是：当病人坚持要求时，它们还能否说“不”。

新闻来源

• EurekAlert! · Ben-Gurion University of the Negev: Does AI know when to say “no”? (20 Oct 2026)
• Annals of Internal Medicine · Ariel, Laban and Endelman study on large language models and low-value care (20 Oct 2026)

✳更多文章 →