Across 300 simulated consultations judged by an independent panel of 20 physicians, it came out level with primary care doctors on history-taking, diagnosis, management and communication — the examination was the one axis where it pulled ahead. No real patients took part.
AMIE, the experimental medical AI from Google Research and Google DeepMind, now works over live video, the two labs said this week, in what they bill as the first demonstration of expert-level AI performance in that kind of encounter. It runs on Gemini and Project Astra, and splits a single consultation across three agents working in parallel: a talker that carries the spoken exchange with the patient, a planner that keeps revising the differential diagnosis in the background, and a perception agent reading the audio and video for clinically meaningful cues the patient never puts into words.
That claim rests on staged visits rather than any real clinic. The trial ran 100 scenarios across five body systems — cardiopulmonary, abdominal, ear-nose-and-throat, neurological and psychiatric, and musculoskeletal — with 15 trained actors carrying out 300 standardized consultations, against 10 board-certified primary care physicians consulting through the same video interface. A separate panel of 20 experienced primary care physicians scored every encounter. On history-taking thoroughness, diagnostic accuracy, appropriateness of the management plan and communication quality, that panel rated AMIE level with the doctors rather than above them. It scored significantly higher than both comparison groups on one axis only: proactively walking the actor through examination maneuvers on camera. It also matched or beat the text-based version of itself on every dimension, and the actors preferred the video format to text chat.
The limits are the researchers' own. Everything ran with professional actors in simulated settings, confined to conditions that can be convincingly acted, and automated evaluation still turned up occasional perceptual and reasoning errors along with intermittent technical faults that broke the flow of the conversation. AMIE has not moved beyond the research stage, and the team calls studies with real patients and real conditions the essential next step.