Israeli Researchers Find AI Recommends Unnecessary Tests in Specific Cases
Translated & summarized from Mignews by baba
Israeli scientists from Ben-Gurion University, Clalit Health Services, and Rabin Medical Center studied AI's medical recommendation capabilities. Published in Annals of Internal Medicine on October 20, 2026, the research found AI generally avoids unnecessary procedures but struggles with ambiguous cases. AI incorrectly recommended unnecessary interventions in 5.4% of non-treatment scenarios and failed to recommend necessary care in 0.6% of cases.
The story in 5 lines · by baba
- Israeli scientists found AI generally recommends against unnecessary medical procedures.
- AI chatbots struggled with ambiguous medical situations, recommending unnecessary tests in 5.4% of cases.
- When medical help was needed, AI mistakenly advised against it in only 0.6% of responses.
- AI most frequently erred in recommending ECGs for low-risk adults, suggesting them unnecessarily in 28.3% of cases.
- The study involved researchers from Ben-Gurion University, Clalit, and Rabin Medical Center.
Israeli scientists have investigated the ability of popular AI chatbots to distinguish between necessary and unnecessary medical procedures. The study, conducted by researchers from Ben-Gurion University of the Negev, Clalit Health Services, and Rabin Medical Center, found that while AI generally avoids recommending superfluous treatments, it struggles with ambiguous medical scenarios. The findings were published on October 20, 2026, in the journal Annals of Internal Medicine.
Researchers tested six language models from Google, Anthropic, and OpenAI, analyzing thousands of responses to hundreds of clinical cases. In situations where no examination or treatment was needed, the AI models incorrectly suggested unnecessary interventions in 5.4% of cases. Conversely, when medical help was required, the models mistakenly advised against it in only 0.6% of responses. In eight out of fifteen tested medical recommendations, the AI did not suggest any unnecessary procedures at all.
The most frequent errors occurred when assessing the need for an electrocardiogram (ECG) in adults with a low risk of cardiovascular disease. In 28.3% of responses, the AI models recommended the test without sufficient indication.
Mentioned
Not the same event — other stories that share this one’s people, places, or theme: background, reactions, and follow-ups.
