Medical AI evaluation has a language gap, a new review finds.

An October 12 review maps where clinical language-model evidence exists. Speaking a language fluently does not establish safety in that setting.

Dean Ariel and Guy Laban reviewed 806 studies covering 62 non-English languages. Their Communications Health paper, published October 12, 2026, finds evaluation concentrated in a few well-resourced languages.

Which evidence was counted?

The review searched PubMed, Scopus and Embase for January 2022–February 2026. It excluded regional-only repositories and preprints, so its map is not a census of all research everywhere.

A proposal, not an approved standard

The authors propose LINGUA to improve language-specific reporting. They say it still needs prospective validation. The review maps evidence gaps; it does not prove that every non-English response is unsafe.

What to ask before relying on a result

Our interpretation: ask which language, task and local context were evaluated. A translated screen or fluent answer is not the same evidence as a test of that clinical workflow. This article reports research, not advice for diagnosing or treating a person.

Published October 14, 2026. Paper announcement: October 12; literature search window ends February 2026. We did not reproduce the review or test a medical AI.

Sources: Communications Health.