Google’s Articulate Medical Intelligence Explorer (AMIE) has successfully evolved from text-based dialogue into a real-time clinical video consultation system, matching or exceeding human primary care physicians across several core competencies in a major 2026 study. Developed by Google Research and Google DeepMind, the upgraded AMIE Video system integrates real-time audio-visual perception, allowing it to interpret non-verbal cues and guide physical examinations just like an expert clinician. [1, 2]

AMIE Video functions through a highly coordinated multi-agent architecture built on Gemini 3 Flash and 3.1 Pro alongside Project Astra. The system features three asynchronously working agents: [1, 3]

• Dialogue Agent (The Talker): Conducts empathetic, real-time audio conversations with low latency. [3, 4, 5]
• Management Reasoning Agent (The Thinker): Continuously refines clinical reasoning, updates potential differential diagnoses, and cross-references authoritative medical guidelines. [4, 5]
• Audio-Visual Perception Agent (The Observer): Constantly analyzes live video streams to look for visible signs of distress, notice physiological symptoms (like a cough), and observe patient movements. [5, 6]

To rigorously test the system, researchers conducted a large-scale, randomized Objective Structured Clinical Examination (OSCE) study. [7]

• Participants: 15 trained professional patient actors, 10 US board-certified primary care physicians (PCPs), and 20 clinical evaluators.
• Scale: 300 live, synchronized clinical video consultations spanning 100 diverse clinical scenarios.
• Scope: Broad medical cases covering 5 body systems, including cardiopulmonary, abdominal, HEENT (head, eyes, ears, nose, throat), neurological/psychiatric, and musculoskeletal conditions.
• The Three Arms: The study directly compared AMIE (Video) against a baseline AMIE (Text) interface and human PCPs (Video) using the exact same interface. [3, 5, 7]

Independent clinical evaluators rated AMIE Video highly across crucial medical parameters, noting that it performed on par with or better than human doctors in several areas: [1, 8]

+------------------------------+-----------------------------------------------------------+

| Clinical Evaluation Axis     | AMIE Video Performance vs. Primary Care Physicians (PCPs) |
+------------------------------+-----------------------------------------------------------+

| History-Taking Thoroughness | Rated favorably, demonstrating exhaustive information gaps |
+------------------------------+-----------------------------------------------------------+

| Diagnostic Accuracy          | High accuracy, aligning strongly with clinical benchmarks |
+------------------------------+-----------------------------------------------------------+

| Management Plan Alignment    | Strictly followed established clinical guidelines         |
+------------------------------+-----------------------------------------------------------+

| Modality Preference          | Patient actors strongly preferred Video over Text Chat    |
+------------------------------+-----------------------------------------------------------+

During the virtual sessions, AMIE demonstrated the ability to guide physical exams by giving precise, real-time instructions, such as asking patients to turn their heads slowly, make a tight fist, or touch their thumb to their index finger to check motor function. It successfully processed verbal feedback regarding pain location (e.g., lower back pain) while providing reassurance about full recovery journeys. [6, 10]

While AMIE matched clinical requirements, the study highlighted clear areas where human doctors still maintain an edge and where the AI must improve: [1, 9]

• Rapport and Trust: Human PCPs still outperform the AI when it comes to relationship building, establishing patient partnership, and nuanced emotional support. [9]
• Perceptual Fluidity: Occasional latency gaps and conversational bottlenecks remain a challenge for the AI system. [9]
• Research Prototype Only: Google emphasizes that AMIE remains strictly an experimental research platform. Substantial real-world trials, safety updates, and regulatory evaluations are mandatory before it can ever be deployed in actual healthcare workflows. [1, 11, 12]