- 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% (0.6–12.2%); false restraint 0.6% (0.1–1.2%)
An Annals benchmark study tested six commercial large language models from three companies on 225 clinician-style scenarios built from 15 Choosing Wisely recommendations, generating 4,050 responses. In each set, restraint was appropriate in some scenarios and action in others.
Overall the models recommended overuse in 5.4% of responses and wrongly withheld an appropriate action in 0.6%. Overuse was concentrated: thyroid ultrasound for an incidental nodule under 1 cm in 42.8% of responses, and a routine ECG in low-risk adults in 28.3%. Most other recommendations were handled well.
Clinicians increasingly check decisions with these tools. The overall picture is reassuring, but the errors cluster in exactly the low-value tests clinicians already over-order, so a tool may reinforce a habit rather than correct it.
Use such tools as a second opinion, not a decision, and be most sceptical when one suggests a test.
- Treat a chatbot's suggestion to order a test with more scepticism than a suggestion to hold off
- Do not order thyroid ultrasound for an incidental nodule under 1 cm without a specific indication
- Do not order a routine ECG in an asymptomatic low-risk adult
- Note that this was a benchmark with single prompts, not real-world use
Why it matters
A decision aid can entrench overuse exactly where it is already most common.
Don't overread it
A benchmark with single prompts cannot show how these tools behave in a real consultation.
The statistics, in plain English
The overall overuse rate of 5.4% has a wide interval (0.6% to 12.2%) because it varied so much between recommendations; the average hides two clear trouble spots.
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.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for top clinical updates, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free