Chatbots mostly practise restraint — except for thyroid nodules and screening ECGs

Design

Benchmark evaluation of six commercial large language models

Population

225 scenarios from 15 Choosing Wisely recommendations; 4,050 responses

Primary outcome

Overuse rate and false restraint rate

Effect

Overuse 5.4% (95% CI 0.6–12.2%); false restraint 0.6%; thyroid nodule ultrasound overuse 42.8%

Researchers turned 15 Choosing Wisely recommendations into 225 clinical scenarios — in two-thirds restraint was right, in one-third action was — and put them, framed as a clinician's question, to six commercial large language models from three companies, three times each. That produced 4,050 responses, classified by an AI judge with physician adjudication.

Overall, models recommended overuse in 5.4% of responses (95% CI 0.6–12.2%) and wrongly withheld appropriate action in 0.6%. But overuse was concentrated: thyroid ultrasound for an incidental subcentimetre nodule was endorsed in 42.8% of responses and a routine ECG in a low-risk adult in 28.3%. Seven of the 15 recommendations drew any overuse at all.

The pattern matters more than the average. The two areas where models erred are ones where low-value testing is already common, so a model tuned on how medicine is practised may reproduce its habits. This was a benchmark with single prompts, not real use, and models change quickly.

• Be sceptical when an AI tool suggests thyroid ultrasound for an incidental nodule under 1 cm.
• Check any AI suggestion of a screening ECG in a low-risk, asymptomatic adult.
• Models rarely withheld appropriate care in this test, so their errors leaned towards doing more.
• Treat AI output as a prompt to think, not a decision, especially for tests with known overuse.

Why it matters

AI tools appear to carry medicine's existing overuse habits with them, in exactly the places where low-value testing is already common.

Don't overread it

A benchmark of single prompts says little about how these tools behave in real conversations or in later versions.

The statistics, in plain English

The overall overuse rate of 5.4% has a wide interval (0.6–12.2%) because it varies so much between scenarios. The averages hide that a few recommendations account for most of the errors.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.