Can AI Give Doctors a Useful Second Opinion?
Eight AI models played a panel of specialists on 190 hard published cases. They reasoned well, but the final call was the weak spot.
In one paragraph
A useful medical second opinion has to stick to the facts, set priorities, plan next steps and stay safe, not just name an illness. This preprint built a benchmark from 300 resolved NEJM case records. Each AI model chose up to five specialist roles and ran a panel review, and an AI judge tuned against six physicians scored 3,040 final reports. Reasoning was usually rated highly, but final conclusions were the weak spot, and which model was used mattered more than the report format.
How were the AI second opinions tested?
The authors used 300 published NEJM case records whose final diagnosis is known: 110 for physician ratings and judge calibration, and 190 held out for the main test. The diagnosis was hidden from the models.
For each case, a model chose one to five clinical roles, wrote an independent opinion for each, then ran a consilium: every role voted on a shared report and a moderator revised it, for at most five rounds, before a final second opinion was written. Eight models each ran the whole pipeline alone, once with reports written in open prose and once with structured reports in fixed sections. An AI judge, calibrated against ratings from six physicians, scored each report from 0 to 3 on eight criteria, including factual grounding, differential diagnosis, management and safety.
What did the AI panels do well and badly?
Reasoning was the strong part. 90.4% of reports received the top score for reasoning coherence, and problem prioritization and differential diagnosis were also near the top of the scale.
The final conclusion was the weak part: 62.8% of final-conclusion scores were 0 or 1. Safety was the second weakest criterion.
Which model was used mattered more than the format. The gap between the best and worst model's final reports was 0.463 points for open-prose reports and 0.572 for structured ones, larger than any within-model format difference. Structured reports scored higher in seven of eight models, mainly through better management plans (2.740 vs. 2.436). The full panel's report scored higher than a single primary-role opinion on every shared criterion, and panels reached consensus in 97.8% of reports, although longer reviews did not score higher.
Why does the weak final conclusion matter?
A second opinion is judged by more than whether it names the right illness: it can change the plan, flag missing information and highlight risks. This benchmark measures those parts separately, so it can show where a fluent answer falls short.
The pattern it finds is that the reasoning is usually sound, while the confident final line is the weakest part. Studies cited in the paper found that biased AI advice can pull clinicians toward wrong decisions, so the final conclusion is the part that most needs checking.
The authors present the benchmark as a way to compare future models, prompting strategies and panel designs. It does not change clinical practice on its own.
What are the limits of this AI second-opinion benchmark?
Read before you quote it
This is a preprint that has not been peer reviewed, and the authors list its limits:
- No patients or clinicians. Only written reports were scored, not effects on clinicians' decisions or patient outcomes.
- Unusual cases. NEJM case records favour rare, difficult cases, not routine practice.
- Possible memorization. The public cases may be in the models' training data.
- An AI judge. It was calibrated on a limited physician-rated sample, and GPT-OSS-120B also judged its own outputs.
- No compute-matched comparison. The panel used more calls and tokens than a single opinion.
The specialist roles are prompted perspectives, not validated specialists.
- It does not show any effect on clinicians' decisions or on patient outcomes.
- It does not show that the specialist roles are real expertise: they are prompted perspectives.
- NEJM case records overrepresent unusual, difficult cases and do not reflect routine practice.
- The public cases may have appeared in the models' training data.
- Scores come from an AI judge calibrated on a limited physician-rated sample, and GPT-OSS-120B also judged its own outputs.
- The panel used more calls and tokens than a single opinion, with no compute-matched comparison.
Takeaways
- saysThe AI panels showed all the working. The final call was the weak spot.
- saysWhich model you pick mattered most. The report's format mattered far less.
- saysAn AI judge scored unusual published cases, so real patients and doctors are still untested.
- explainsA consilium is a panel of specialists who review a case together and agree on one written opinion.
- interpretsA fair reading is that AI second opinions need the closest check on their confident final line.
Ask the paper
In PaperSays you can ask this paper anything, by text or by saying “Hey Aiden”. Three questions listeners asked:
What is a differential diagnosis?
The short list of illnesses that could explain the symptoms, ranked from most to least likely. Hear the full answer in the app →
Who scored the AI reports?
An AI judge, calibrated against ratings from six physicians on a separate set of cases. Ask in the app →
Did structured reports help?
Mainly for the management plan; which model was used mattered much more. Ask in the app →
Briefing reviewed by an editor before publication. The paper belongs to its authors; this page summarises it and links to the original. Reviewed by Editor. Paper licence: CC BY 4.0. Corrections: support@papersays.com.