Google tests AMIE on video calls: on a par with primary care doctors across 300 simulated consultations

On 11 August 2026 Google Research and Google DeepMind published on their official blog the results of AMIE (Video), the extension of the AMIE medical research system to real-time video consultations (Google Research, official blog, 11 Aug 2026). The corresponding preprint, “Towards Expert-level Medical AI for Real-time Video Consultations”, was posted to arXiv on 10 August 2026 (arXiv 2608.09861) and lists around 40 researchers, among them Mahvish Nagda, Jihyeon Lee, Matthew Thompson and Chunjong Park. AMIE (Video) is built on Gemini and Project Astra and uses an architecture in which three agents work in parallel: **Talker** runs the conversation with the patient, **Planner** reasons in the background and updates the differential diagnosis, **Perception** continuously analyses the audio-video stream to catch non-verbal cues and physical findings (Google Research, official blog, 11 Aug 2026).

The study was run as a randomised OSCE: 100 clinical scenarios, 300 live consultations, 15 trained patient-actors and three comparison arms – AMIE (Video), AMIE (Text) and general practitioners on a video call. The human arm included 10 board-certified primary care physicians, assessed by an independent panel of 20 specialist physicians, for a total of 30 primary care doctors involved (Google Research, official blog, 11 Aug 2026). It is the method used to examine medical students: constructed scenarios, actors trained to play out a clinical picture. Not one of the 300 consultations involved a person with a real health problem.

The clinical raters judged AMIE (Video) on a par with the primary care doctors on history taking, diagnostic accuracy, appropriateness of management and quality of communication, giving it significantly higher scores than both the physicians and the text-only version on eliciting physical signs and guiding examination manoeuvres remotely; against the text version, AMIE (Video) matched or beat the result in every category measured (Google Research, official blog, 11 Aug 2026). In the head-to-head with the doctors, the patient-actors preferred AMIE's way of assessing and explaining conditions, while they preferred the human doctors for building rapport and the therapeutic alliance (Google Research, official blog, 11 Aug 2026). The numerical scores and the confidence intervals were not disclosed in the blog post, but they are in the preprint.

Google stresses that AMIE remains a research system, neither approved nor available for clinical use, and that studies with real patients are needed before drawing conclusions about its use in care (Google Research, official blog, 11 Aug 2026). The preprint does not appear to have been peer reviewed yet, unlike earlier work on the same research line published in scientific journals. The preprint flags occasional perception and reasoning errors, as well as limitations in anatomical precision and in subtle affective nuance (Nagda et al., arXiv 2608.09861). It is not public whether or when a clinical study with real patients will begin, nor whether AMIE (Video) will be turned into a commercial product.