A study published in Annals of Internal Medicine on October 20, 2026 found that leading large language models correctly identify redundant medical interventions in most clinical cases.
Researchers from Ben-Gurion University of the Negev and Clalit Health Services evaluated six large language models from Google, Anthropic, and OpenAI against Choosing Wisely recommendations.
Large language models recommended unnecessary EKGs in 28.3% of low-risk patient cases, primarily when patients reported palpitations, in an October 20, 2026 study.
Large language models recommended unnecessary thyroid ultrasounds in 42.8% of incidental nodule cases, mirroring clinical disagreements between the American Thyroid Association and American College of Radiology.
According to an October 20, 2026 study, large language models inappropriately recommended unnecessary medical tests or treatments in 5.4% of explicitly unneeded cases and recommended withholding care in 0.6% of medically indicated cases.
Large language models recommended zero unnecessary procedures across eight of 15 Choosing Wisely recommendations evaluated in an October 20, 2026 study.
Recommendations for identical ambiguous clinical scenarios varied widely between large language models, with one model suggesting unnecessary tests in under 10% of cases and another in 80%.
Researchers in an October 20, 2026 study evaluated artificial intelligence models across hundreds of clinical scenarios, analyzing thousands of total model responses.