Source study found

Story checked

New study reveals AI chatbots recognize unnecessary medical tests, but stumble where human doctors disagree (opens in a new tab)

medicalxpress.com · 2026-10-20

Short answerEvidenceSource

Short answer

Mixed

Mixed.

2 claims go further than the study. 2 other points were not covered by the paper.

3 supported
2 overstated
2 not covered

Checked against the study summary. The full text wasn't available, so some details couldn't be settled either way.

Share this check

Follow the evidence trail

The story

New study reveals AI chatbots recognize unnecessary medical tests, but stumble where human doctors disagree

medicalxpress.com · 2026-10-20

The story’s checkable claims.

Read the original story

NewsLink checks it

Mixed

Two of seven claims overstate the study. Three of seven check out. Two claims the study doesn't address.

3 supported
2 overstated
2 not covered

Open claim evidence

The source study

Overuse and False Restraint in Large Language Model Recommendations for Choosing Wisely Scenarios

Annals of Internal Medicine · 2026

Open source record Open in Tessa

Then inspect each claim

Evidence layer

Claim by claim

Each claim gets a verdict. Expand it to see the evidence directly below.

Reading mode

Scan verdicts. Open evidence only when needed.

Browse by verdict

7 claims in this story

Showing all 7 claimsChoose a verdict to focus the list.

Claim 1 of 7OverstatedA new joint study from Ben-Gurion University of the Negev and Clalit Health Services found that leading large language models identify when medical interventions are redundant in the vast majority of cases, but stumble mainly in gray areas where medical consensus is divided.View evidenceHide evidence

As statedvast majority of cases

Why this verdict

The abstract-level profile supports that six commercial LLMs had low overall overuse and false-restraint rates, so they often recommended restraint appropriately. It also supports that overuse was concentrated in two recommendation topics. However, the headline framing that the models mainly stumble in 'gray areas where medical consensus is divided' is not established in the supplied paper profile; the profile identifies specific Choosing Wisely topics with high overuse but does not show that these are areas of divided medical consensus. As a headline claim, this outruns the abstract-level evidence.

Study evidence

Overall LLM-recommended overuse rate was 5.4% (95% CI, 0.6% to 12.2%).5.4% (95% CI, 0.6% to 12.2%)

“DESIGN: Benchmark evaluation.”

Study evidence

Two Choosing Wisely recommendations accounted for most LLM-recommended overuse: thyroid ultrasound for an incidental subcentimeter nodule and routine electrocardiography in low-risk adults.Thyroid ultrasound: 42.8% of LLM responses recommended overuse; routine ECG: 28.3% of LLM responses recommended overuse.

“Two recommendations accounted for most LLM-recommended overuse: thyroid ultrasound for an incidental subcentimeter nodule ... and routine electrocardiography in low-risk adults”

Claim 2 of 7OverstatedMost AI overtesting errors clustered around routine EKGs in low-risk patients and thyroid ultrasounds for incidental nodules, with especially high unnecessary-test rates in those ambiguous scenarios.View evidenceHide evidence

As stated28.3% for routine EKGs and 42.8% for thyroid ultrasounds

Why this verdict

The paper profile supports the main quantitative claim that most overuse was concentrated in thyroid ultrasound for an incidental subcentimeter nodule and routine ECG in low-risk adults, with rates of 42.8% and 28.3%, respectively. However, the story additionally characterizes these as 'ambiguous' or gray-area scenarios, which is not established in the abstract-level profile. The error concentration is supported; the ambiguity/consensus interpretation is not.

Study evidence

Overall LLM-recommended overuse rate was 5.4% (95% CI, 0.6% to 12.2%).5.4% (95% CI, 0.6% to 12.2%)

“DESIGN: Benchmark evaluation.”

Study evidence

“Two recommendations accounted for most LLM-recommended overuse: thyroid ultrasound for an incidental subcentimeter nodule ... and routine electrocardiography in low-risk adults”

Claim 3 of 7Not coveredThe study was published today in Annals of Internal Medicine and was led by Dr. Dean Ariel, Dr. Guy Laban, and Dr. Stav Endelman.View evidenceHide evidence

Why this verdict

The supplied abstract-level paper profile does not provide the journal publication timing, 'published today' status, or named lead authors. These bibliographic/authorship details may be true, but they are not verifiable from the supplied profile.

Claim 4 of 7Not coveredThe article says one future research question is whether an AI assistant will maintain appropriate clinical restraint when an anxious or insistent patient pushes back.View evidenceHide evidence

Why this verdict

The abstract-level profile notes limitations involving single prompts, lack of follow-up, benchmark rather than real-world use, and uncertain consumer-facing generalizability. It does not specifically identify anxious or insistent patient pushback as a future research question. This may appear in the full article discussion, but it is not verifiable at the supplied abstract depth.

Study evidence

Benchmark construction: 225 vignettes were derived from 15 Choosing Wisely recommendations, with an explicit balance of scenarios where restraint or action was appropriate.

“Fifteen Choosing Wisely recommendations were used to develop 225 scenarios”

Claim 5 of 7SupportedResearchers evaluated six leading LLMs from Google, Anthropic, and OpenAI across hundreds of clinical scenarios built around Choosing Wisely recommendations, analyzing thousands of total model responses.View evidenceHide evidence

As statedhundreds of clinical scenarios; thousands of total model responses

Why this verdict

The profile states that six commercial LLMs from OpenAI, Anthropic, and Google were queried through production APIs on 225 Choosing Wisely-derived scenarios, with three runs per model producing 4,050 total responses. This supports the claim of six models, hundreds of scenarios, and thousands of responses.

Study evidence

Overall LLM-recommended overuse rate was 5.4% (95% CI, 0.6% to 12.2%).5.4% (95% CI, 0.6% to 12.2%)

“DESIGN: Benchmark evaluation.”

Study evidence

Benchmark construction: 225 vignettes were derived from 15 Choosing Wisely recommendations, with an explicit balance of scenarios where restraint or action was appropriate.

“Fifteen Choosing Wisely recommendations were used to develop 225 scenarios”

Claim 6 of 7SupportedWhen an intervention was medically indicated, the models mistakenly recommended withholding care in only 0.6% of cases, and when an intervention was explicitly unneeded, they inappropriately recommended testing or treatment in 5.4% of cases.View evidenceHide evidence

As stated0.6% and 5.4%

Why this verdict

The profile reports the same co-primary outcomes: false restraint, defined as withholding appropriate action, was 0.6%, and overuse, defined as endorsing action when restraint was appropriate, was 5.4%.

Study evidence

Overall LLM-recommended overuse rate was 5.4% (95% CI, 0.6% to 12.2%).5.4% (95% CI, 0.6% to 12.2%)

“DESIGN: Benchmark evaluation.”

Claim 7 of 7SupportedIn eight of the 15 Choosing Wisely recommendations, the LLMs did not recommend a single unnecessary procedure.View evidenceHide evidence

As statedeight of the 15 recommendations

Why this verdict

The profile explicitly reports that overuse was absent for 8 of the 15 Choosing Wisely recommendations.

Study evidence

Overall LLM-recommended overuse rate was 5.4% (95% CI, 0.6% to 12.2%).5.4% (95% CI, 0.6% to 12.2%)

“DESIGN: Benchmark evaluation.”

Study evidence

“Two recommendations accounted for most LLM-recommended overuse: thyroid ultrasound for an incidental subcentimeter nodule ... and routine electrocardiography in low-risk adults”

Then look for missing context

Context layer

What the story left out

Important study details the story did not include.

False restraint was absent for 11 of 15 recommendations and highest for PSA screening at 3.3%.
The story reports the overall false-restraint rate but does not mention the recommendation-level pattern for false restraint, including PSA screening as the highest false-restraint topic.
From benchmark evaluation; Stratified subgroup analysis of benchmark LLM responses
Important limitation: the study was an in-silico benchmark using constructed scenarios, not a real-world clinical deployment or clinician-in-the-loop evaluation.
The story presentation does not indicate that the evidence comes from a constructed benchmark rather than real-world use, which is material to interpreting how far the findings generalize.
From benchmark evaluation; Scenario-based benchmark with automated LLM judging and physician adjudication
Important limitation: prompts were single-turn clinician-perspective prompts without follow-up or interactive clarification.
The story mentions a possible future question about patient pushback, but it does not clearly report the paper limitation that the benchmark used single prompts without follow-up. This limitation matters for interpreting model behavior in interactive clinical settings.
From benchmark evaluation; Scenario-based benchmark with automated LLM judging and physician adjudication
Important limitation: responses were classified by an LLM judge with physician adjudication, creating possible classification or adjudication limitations.
The story caveats provided do not mention the LLM-as-judge classification approach or the limited abstract-level detail about adjudication procedures.
From benchmark evaluation; Scenario-based benchmark with automated LLM judging and physician adjudication
Important limitation: findings may not extend to consumer-facing use or other prompt framings.
The story caveats focus on disagreement among doctors and model variability, but do not acknowledge the paper's stated uncertainty about generalizability to consumer-facing use or different prompting contexts.
From benchmark evaluation; Scenario-based benchmark with automated LLM judging and physician adjudication

3 things the story did carry across

Core benchmark design: six commercial LLMs were evaluated on 225 Choosing Wisely-derived clinical scenarios, with three runs per model and 4,050 total responses.
Primary outcome results: overall overuse was 5.4% and false restraint was 0.6%.
Recommendation-level error concentration: overuse was concentrated in thyroid ultrasound for incidental subcentimeter nodules and routine ECG in low-risk adults; overuse was absent for 8 of 15 recommendations.

Then read the study layer

Study layer

Study at a glance

Scan the study first. Expand only the parts you want to inspect.

Pieces of work

Evidence read

study summary

Lead result

in silico

1Lead resultin silicoBenchmark whether commercial LLMs appropriately recommend restraint vs action on clinical scenarios derived from Choosing Wisely recommendations (quantify overuse and false restraint rates overall and by recommendation/LLM).benchmark evaluationExpand

In plain English

Benchmark evaluation of six commercial large language models (OpenAI, Anthropic, Google production APIs) responding to 225 clinical scenarios derived from 15 Choosing Wisely recommendations (per recommendation: 10 scenarios where restraint was appropriate; 5 where action was appropriate). Models were run 3 times each (4050 total responses) using a clinician-perspective prompt. Co-primary outcomes were overuse rate (endorsing action when restraint appropriate) and false restraint rate (withholding appropriate action). Overall overuse was 5.4% (95% CI, 0.6% to 12.2%); false restraint was 0.6% (95% CI, 0.1% to 1.2%). Overuse was concentrated in two recommendations (thyroid ultrasound for incidental subcentimeter nodule: 42.8% of LLM responses; routine ECG in low-risk adults: 28.3%); overuse was absent for 8 of 15 recommendations and low (0.6%–3.3%) across five other recommendations. False restraint was absent for 11 of 15 recommendations and highest for PSA screening (3.3%). Limitations include single prompts without follow-up, benchmark (not real-world) setting, response classification performed by an LLM judge with physician adjudication, and uncertain generalizability to consumer-facing use.

Key findings

Overall LLM-recommended overuse rate was 5.4% (95% CI, 0.6% to 12.2%).5.4% (95% CI, 0.6% to 12.2%)
Overall false restraint rate (withholding appropriate action) was 0.6% (95% CI, 0.1% to 1.2%).0.6% (95% CI, 0.1% to 1.2%)

“DESIGN: Benchmark evaluation.”

What this piece can’t prove

Single prompts without follow-up or interactive clarification.
Benchmark (constructed scenarios) rather than real-world deployment or clinician-in-the-loop evaluation.
Response classification performed by an LLM judge with physician adjudication (potential for classification bias).
Findings may not extend to consumer-facing use or other prompt framings.
No human participants; in-silico setting only.

2in silicoIdentify which specific Choosing Wisely recommendations disproportionately drive LLM-recommended overuse and false restraint (error concentration analysis).Stratified subgroup analysis of benchmark LLM responsesExpand

In plain English

Secondary analysis of error concentration by Choosing Wisely topic within a benchmark of commercial LLM responses found that two recommendations (thyroid ultrasound for an incidental subcentimeter nodule; routine electrocardiography in low-risk adults) accounted for most LLM-recommended overuse, while false restraint was rare overall and highest for prostate-specific antigen (PSA) screening.

Key findings

Two Choosing Wisely recommendations accounted for most LLM-recommended overuse: thyroid ultrasound for an incidental subcentimeter nodule and routine electrocardiography in low-risk adults.Thyroid ultrasound: 42.8% of LLM responses recommended overuse; routine ECG: 28.3% of LLM responses recommended overuse.
Overall LLM-recommended overuse was uncommon across the full set of recommendations.Overall overuse rate 5.4% (95% CI, 0.6% to 12.2%).

“Two recommendations accounted for most LLM-recommended overuse: thyroid ultrasound for an incidental subcentimeter nodule ... and routine electrocardiography in low-risk adults”

What this piece can’t prove

Single prompts without follow-up; results may differ with iterative or interactive prompting.
Benchmark (in-silico) evaluation rather than real-world clinical use; findings may not generalize to deployed or consumer-facing settings.
Response classification conducted by an LLM judge with physician adjudication; potential for misclassification or adjudication bias.

1 further detail could not be confirmed from the summary.

3in silicoDescribe/characterize the benchmark setup (scenario construction, prompt/run structure, and response classification approach including LLM-judge with physician adjudication).Scenario-based benchmark with automated LLM judging and physician adjudicationExpand

In plain English

The study constructed a scenario-based benchmark using 15 Choosing Wisely recommendations to produce 225 vignettes (per recommendation: 10 where restraint was appropriate, 5 where action was appropriate), queried six commercial LLMs via production APIs (OpenAI, Anthropic, Google) with a single clinician-perspective prompt (3 runs per model, total 4,050 responses), and classified responses using an automated LLM-as-judge workflow with physician adjudication.

Key findings

Benchmark construction: 225 vignettes were derived from 15 Choosing Wisely recommendations, with an explicit balance of scenarios where restraint or action was appropriate.
Prompt and run structure: a single clinician-perspective prompt was used to query six commercial LLMs (production APIs) with three runs per model, producing 4,050 responses.

“Fifteen Choosing Wisely recommendations were used to develop 225 scenarios”

What this piece can’t prove

Single prompts without follow-up; benchmark rather than real-world use.
Responses classified by an LLM judge with physician adjudication; adjudication procedures and reliability not detailed in abstract.
Findings may not extend to consumer-facing use.

Finally, the search trail

Method layer