1 hour ago
18
Google’s medical AI system, AMIE, has graduated from typing diagnoses in a chat window to actually watching patients on video and guiding physical exams in real time. What started as a text-based research project two years ago is now conducting supervised telehealth consultations that rival the diagnostic accuracy of primary care physicians.
A feasibility study at Beth Israel Deaconess Medical Center put AMIE through its paces with 100 real patients in an ambulatory clinic setting. The system hit a 90% accuracy rate on differential diagnoses, matching the performance of the human doctors it was benchmarked against. Zero safety interventions were required during the entire deployment.
From chat bot to co-clinician
AMIE, short for Articulate Medical Intelligence Explorer, first appeared in a January 2024 Nature paper as a large language model designed for diagnostic dialogue. Think of it as a system trained not just to answer medical questions, but to ask the right ones, systematically working through a patient’s symptoms the way an experienced clinician would during an intake interview.
Built on Google’s Gemini models and its Project Astra framework, AMIE now processes live audio and video feeds during simulated telehealth encounters. Instead of just reading text inputs, it can observe a patient on camera and provide real-time instructions for physical examinations.
In April 2026, Google tested these co-clinician capabilities across 120 hypothetical telehealth scenarios. The system successfully guided physical exams while operating under the supervision of licensed physicians.
Across broader evaluations, AMIE matched or exceeded primary care physician performance in 68 out of 140 consultation skill domains.
The BIDMC study in detail
The March 2026 prospective study at Beth Israel Deaconess Medical Center was the first time AMIE handled pre-visit history-taking with actual patients rather than actors. One hundred patients participated in a supervised ambulatory clinic setting, with AMIE conducting the initial interview before patients saw their physician.
Throughout the entire study, clinicians supervising the system never had to step in to correct a dangerous suggestion or redirect a conversation.
A separate June 2026 publication in Nature pushed the evaluation further, testing AMIE’s ability to manage patients across multiple visits rather than single encounters. The system met or surpassed physician standards in care planning for longitudinal disease management, using patient actors to simulate ongoing treatment relationships.
Supervised, not autonomous
Google has been emphatic on one point across every publication: AMIE is not designed to practice medicine independently. Every deployment scenario involves physician oversight, and the company has consistently framed the technology as a tool that augments clinical workflows rather than replacing them.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·