Six AI models rarely recommended excess care, but two test decisions drove most errors

Six commercial AI models usually distinguished unnecessary medical interventions from cases in which action was warranted in a new Choosing Wisely benchmark. Their mistakes were concentrated: thyroid ultrasound for incidental small nodules and routine electrocardiograms for low-risk adults accounted for most recommendations for excess care. That concentration is more informative than the reassuring overall average.[3]

They built examples from Choosing Wisely, an initiative that asks clinical organisations to identify tests and treatments that are often ordered without enough benefit. Crucially, the researchers constructed neighbouring scenarios in which a test was appropriate. A system that always said no would fail too. The objective was not generic diagnostic brilliance but discrimination between low-value action and a medically justified exception.[3]

Counting the right mistakes

Fifteen guideline recommendations produced 225 clinical vignettes. For each recommendation, ten cases called for restraint and five called for action. Six commercial LLMs answered each example three times through production APIs, creating 4,050 text responses. The team then classified whether each answer endorsed too much testing or incorrectly withheld indicated care. An automated judge helped label answers, with physician adjudication; the classification procedure adds its own potential uncertainty.[3]

Across the evaluated restraint scenarios, the mean rate of recommendations for unnecessary care was 5.4 percent, with a broad reported uncertainty interval from 0.6 to 12.2 percent. For scenarios in which action was justified, false restraint was estimated at 0.6 percent. Eight of the fifteen recommendations produced no overuse at all. These numbers are not a combined diagnostic accuracy rate: the two errors use different sets of scenarios, and the average masks enormous variation by topic.[1][2]

The largest pockets of overuse concerned ultrasound for an incidental thyroid nodule smaller than one centimetre, where it appeared in 42.8 percent of responses, and routine electrocardiography for low-risk adults, where it appeared in 28.3 percent. The numbers describe model responses on the particular benchmark examples, not the actual frequency of unnecessary ultrasounds or ECGs in hospitals. Some uncertainty reflects real differences between medical societies over exactly when a finding deserves more investigation.[3]

A model can inherit a profession's argument

Medical guidance is not a single timeless rulebook. Different professional groups weigh false alarms, missed disease, costs and patient anxiety differently. If a model's training material contains several authoritative but inconsistent recommendations, it may sound entirely plausible whichever path it selects. A headline score can conceal this disagreement until the test questions are arranged to expose it. In those situations, explainable reasoning and awareness of guidance differences matter more than fluency.[3]

Consider an incidental small thyroid nodule. Whether imaging is helpful may depend on features that a short vignette cannot capture, as well as the recommendation being applied. A system may reflexively advise a follow-up because its training rewards thoroughness. Conversely, reflexive restraint could become unsafe if a patient has red flags not represented in the example. This is why the study placed positive action cases beside restraint cases rather than scoring compliance with a list of isolated prohibitions.[3]

A benchmark designed to catch both kinds of error

A system that refuses every test would look restrained but fail patients who need care. The study therefore paired 150 restraint scenarios with 75 action scenarios, rather than simply checking whether models repeated a list of prohibitions. False restraint was uncommon in the resulting responses, which is a useful complement to the overuse measure. Both rates apply to these written cases and model configurations, tested from February to March 2026.[3]

The benchmark used one clinician-perspective prompt for each case, with no follow-up conversation and no human participants. It therefore leaves open what happens when symptoms change, a history is incomplete or a user pushes repeatedly for a different answer. Those are specific gaps between the experiment and a consultation, rather than reasons to dismiss the finding.[3]

For developers, the clearest target is the cluster of difficult decisions. A model needs to recognise when professional guidance differs and which case details determine whether a test is useful. Improving those distinctions would do more for this benchmark than simply training it to give a more confident yes or no.

Report an error or suggest a correction