Israeli study finds AI models usually recognize unnecessary medical care

Leading artificial intelligence chatbots generally recognized when medical tests and treatments were unnecessary in an Israeli study, although their recommendations varied considerably in disputed areas of medicine. Researchers from Ben-Gurion University of the Negev and Clalit Health Services also found that the models rarely advised withholding care when an intervention was medically warranted.

The peer-reviewed research appeared in Annals of Internal Medicine, the American College of Physicians’ flagship journal. The team included Prof. Guy Laban of Ben-Gurion University’s Department of Industrial Engineering and Management, Clalit family physician and researcher Dr. Dean Ariel, and Dr. Stav Endelman of Rabin Medical Center-Beilinson Hospital in Petah Tikva.

The researchers examined six leading large language models from Google, Anthropic and OpenAI, analyzing thousands of responses to hundreds of clinical scenarios. Their cases drew on 15 recommendations from Choosing Wisely, an international campaign established in 2012 by the American Board of Internal Medicine Foundation to reduce low-value care. Written from a clinician’s perspective, the scenarios paired situations requiring restraint with closely matched cases in which the same intervention was necessary.

Examples illustrated how a patient’s circumstances change the appropriate response. Acute lower back pain after lifting furniture in a young man did not require imaging, while new nighttime back pain in a woman with a history of breast cancer did. Similarly, antibiotics were unwarranted for a common viral cold but appropriate following a confirmed bacterial strep throat test.

Across cases where an intervention was indicated, the models incorrectly recommended withholding it in 0.6% of responses. Where testing or treatment was explicitly unnecessary, they recommended proceeding in 5.4% of responses. For eight of the 15 recommendations examined, the models suggested no unnecessary procedures.

Most excessive testing recommendations involved two practices with clinical ambiguity. Models endorsed unnecessary electrocardiograms in 28.3% of responses, mainly in cases mentioning palpitations, for which such testing can reasonably form part of an assessment. They recommended unnecessary ultrasounds for small, incidentally discovered nodules in 42.8% of responses. The researchers linked that finding to disagreement between the American Thyroid Association and the American College of Radiology about whether those nodules require imaging.

Laban and Ariel told The Jerusalem Post that unnecessary investigations can reveal harmless findings and lead to anxiety, specialist referrals, biopsies or surgery. They also stressed the danger of excessive restraint when patients need care. Laban said differences between models were especially pronounced in disputed cases: one recommended an unnecessary test less than 10% of the time, while another did so 80% of the time.

The findings leave open whether the systems demonstrate clinical judgment or successfully match established guidance to constructed cases. Testing was conducted only in English. Laban identified another unresolved question: whether a chatbot would maintain appropriate restraint when an anxious or insistent patient challenged its recommendation.

Ariel said further research, including prospective studies, was needed before the findings could be clinically validated and applied in medical practice. He emphasized the need for safeguards and said responsibility ultimately rests with the doctor, with AI providing recommendations. He also questioned whether using AI to avoid tests would necessarily reduce costs.