A second opinion tested across three health systems.

The human problem

Clinicians must apply changing guidance under time and information pressure, with uneven access to specialist support.

The prior constraint

General-purpose clinical assistants may perform differently across languages, guidelines, and health systems, and many evaluations do not include practicing clinicians across countries.

AI’s actual role

GPT-4o supplied information and reasoning support while physicians retained the task and answer.

The documented result

Among 249 physicians, guideline scores increased by 18 percentage points in Kenya, 10.7 in Indonesia, and 7.2 in the Netherlands. These were vignette scores, not patient outcomes.

Why it may matter

The study asks whether assistance transfers across settings while making the geographic differences visible.

Limitations

The tasks were clinical vignettes, not live care.

Guideline-score gains do not establish safety, diagnostic accuracy, or patient benefit.

One model version and study interface may not generalize to other tools or changing models.

Unresolved questions

Which specialties and case types account for errors or gains?

How does assistance affect time, overreliance, and disagreement in real practice?

Do language and local guideline differences change safety?