Language Models Show Restraint on Low-Value Care, Falter in Clinical Gray Areas — Research Library

Key Highlights

Six large language models from Google, Anthropic and OpenAI recommended unnecessary intervention in 5.4% of cases where clinical restraint was warranted and withheld indicated care in 0.6%.
Accuracy fell on two contested questions: unnecessary ultrasound for incidental small thyroid nodules in 42.8% of responses, and unnecessary routine electrocardiograms for low-risk patients in 28.3%.
On identical cases without professional consensus, individual models ranged from under 10% to 80% in recommending unnecessary tests.

Context and Core Dynamics

A study by researchers at Ben-Gurion University of the Negev and Clalit Health Services examined whether large language models, trained on the clinical literature physicians themselves produce, have absorbed clinical restraint. Published in the Annals of Internal Medicine, it tested six models developed by Google, Anthropic and OpenAI on decisions about whether to test or treat a patient, or leave well enough alone.

Results were largely reassuring. The models recommended unnecessary intervention in 5.4% of cases where restraint was called for, and withheld medically indicated care in 0.6%. In eight of the 15 Choosing Wisely recommendations examined, no model proposed an unnecessary procedure.

Degradation concentrated in two gray areas. For incidental small thyroid nodules, models recommended unnecessary ultrasound in 42.8% of responses, a question on which the American Thyroid Association and the American College of Radiology diverge. They recommended unnecessary routine electrocardiograms for low-risk patients in 28.3% of responses, particularly where palpitations were reported. Dean Ariel, a family physician and medical AI researcher at Clalit Health Services (Tel Aviv District), framed the difficulty: "One of the hardest parts of practicing medicine is knowing when to hold back." Performance also diverged sharply where no professional consensus exists; on identical cases, one model recommended an unnecessary test less than 10% of the time and another 80% of the time, Guy Laban of Ben-Gurion University of the Negev reported.

Caregiving Application and Social Impact

The immediate application lies in eldercare, where comorbidity, polypharmacy and incidental imaging findings make low-value testing a standing risk. A decision-support layer that reliably declines to escalate, and flags the narrow set of contested questions where models disagree with each other, could assist home-care nurses, geriatric assessment teams and rehabilitation services without displacing clinical judgement. Because divergence was widest where physicians themselves hesitate, deployment should pair model output with documented local policy on contested screening questions. Laban's next question, whether models retain restraint when anxious patients insist on further testing, is familiar to anyone supporting older adults through diagnostic uncertainty.

Ethics and Compliance Note

Accountability for ordering or withholding a test remains with the treating clinician, and this evidence does not support autonomous test ordering. Wide inter-model variance on consensus-free cases argues for version pinning, documented model provenance in the patient record, and periodic revalidation against local practice. Patient data minimisation and jurisdiction-specific rules on clinical decision support apply, and contested recommendations should reach clinicians as flagged uncertainty rather than silent resolution.