Overview
- The Nature preprint published June 17, 2026, reports a randomized, blinded virtual Objective Structured Clinical Examination that compared AMIE to 21 primary care physicians across 100 multi‑visit scenarios and found AMIE non‑inferior on management reasoning.
- AMIE used Gemini’s long‑context modeling, in‑context retrieval, and structured reasoning to ground recommendations in UK NICE and BMJ Best Practice guidance and in drug formularies, and it scored higher than PCPs on preciseness of treatments and guideline alignment.
- The study introduced RxQA, a pharmacist‑validated multiple‑choice benchmark drawn from US and UK formularies, on which AMIE outperformed clinicians on higher‑difficulty medication questions when both could access external drug information.
- All testing used simulated workflows with patient‑actors and specialist raters, and the authors caution these controlled results do not show readiness for real clinical use because of remaining safety, consistency, liability, and integration challenges.
- Secondary reporting has added disputed numeric details and a multimodal extension that processes images and ECGs, but those claims differ from the Nature manuscript and require separate verification before being treated as established.