Google’s AMIE AI Matches Primary Care Doctors in Simulated Video Consultations

Google Research says its medical AI system, AMIE, has performed on par with primary care physicians in simulated real-time video consultations, showing promising results in history-taking, diagnostic reasoning, treatment planning, communication, and virtual physical examinations.

But Google is also making one thing clear: AMIE is not yet ready to replace doctors or independently treat real patients.

The latest research used professional patient actors rather than actual patients, meaning further clinical studies are still needed before the technology can be considered for real-world medical use.

AMIE Uses Three AI Agents During a Consultation

Instead of relying on a single AI model to handle everything, Google designed AMIE around a multi-agent architecture.

The system divides the medical consultation among three specialized AI agents:

The Talker Agent communicates directly with the patient. It handles the spoken conversation, asks questions, responds naturally, and maintains the flow of the consultation.

The Planner Agent works in the background. It continuously updates possible diagnoses, treatment plans, missing information, and the next clinical questions that should be asked.

The Perception Agent analyzes the video and audio streams for visual and auditory clues. These may include physical signs, movements, breathing patterns, speech characteristics, and other observations that could be medically relevant.

The architecture addresses one of the major technical challenges of conversational medical AI: latency.

Complex medical reasoning can take longer than normal conversation allows. If the AI pauses too long every time it needs to analyze a case, the interaction can feel unnatural.

By separating conversation from deeper reasoning and perception, AMIE's patient-facing agent can continue responding while other agents process clinical information in the background.

AMIE Compared With Real Primary Care Physicians

Google conducted a randomized evaluation involving 15 trained patient actors portraying medical conditions involving five major areas:

• Cardiopulmonary
• Abdominal
• Head, eyes, ears, nose, and throat
• Neurological or psychiatric
• Musculoskeletal

AMIE conducted synchronous video consultations with the actors.

Researchers also tested a text-only version of AMIE and compared both systems with 10 board-certified primary care physicians using the same video consultation interface.

A separate panel of 20 experienced primary care physicians evaluated the consultations using established clinical assessment criteria.

According to Google, evaluators rated video-based AMIE at a level comparable with the physician group in several major areas, including:

• Thoroughness of medical history-taking
• Diagnostic accuracy
• Appropriateness of medical management
• Patient-centered communication

The video version of AMIE also matched or exceeded the text-only version across several measurements.

AI Performed Strongly in Virtual Physical Examinations

One of the more notable findings involved physical examination guidance.

Evaluators reportedly rated AMIE's video system higher than both the physician group and the text-only AI when it came to identifying physical signs and proactively instructing patient actors to perform examination maneuvers through video.

For example, an AI system could ask someone to demonstrate movement, show a particular body area to the camera, perform a basic motor task, or provide another visual clue that could help with clinical reasoning.

The perception system could then place those observations into the wider medical context of the consultation.

This is important because traditional text-based medical AI systems largely depend on symptoms that patients describe verbally.

Video introduces another layer of information.

An AI could potentially observe how a patient moves, speaks, breathes, reacts, or performs an instructed examination.

Patient Actors Preferred Video Over Text

The professional patient actors participating in the study also reportedly preferred the synchronous video experience over text chat.

Google says participants found video easier to use and more effective for explaining their medical concerns.

They also rated AMIE favorably in areas such as empathy, rapport, and confidence in the consultation.

These findings suggest that future healthcare AI may not simply look like a chatbot where patients type symptoms into a box.

Instead, medical AI systems could eventually become multimodal virtual health assistants capable of seeing, hearing, speaking, reasoning, and interacting in real time.

Google Tested AMIE Before Human Evaluation

Before conducting the physician-reviewed study, Google developed an automated testing framework to evaluate specific telehealth capabilities.

The system tested areas involving visual perception, audio signals, clinical reasoning, conversational performance, and virtual physical examinations.

Some tests examined individual tasks, such as recognizing anatomical laterality or possible signs of respiratory distress.

Other simulations involved longer medical conversations.

In one Parkinson's disease scenario, for example, the system could receive information indicating that a patient showed the camera handwriting that had become unusually small and cramped.

These automated evaluations allowed researchers to identify weaknesses and improve AMIE before conducting the actor-based clinical simulation.

However, Google acknowledges that controlled simulations cannot fully reproduce real healthcare environments.

The Biggest Limitation: These Were Not Real Patients

This distinction is critical.

The research does not prove that AMIE can safely diagnose and manage actual patients.

Professional actors cannot reproduce the full unpredictability of real medical cases.

Real patients may have complicated medical histories, multiple diseases, unclear symptoms, poor internet connections, difficult camera angles, environmental distractions, language barriers, emotional distress, or symptoms that cannot easily be simulated.

Some medical conditions were also excluded because actors could not realistically reproduce the necessary physical signs.

Researchers additionally observed occasional perception and reasoning errors, as well as technical interruptions that affected conversational quality.

For these reasons, Google says research involving real patients with actual medical conditions must come next.

Google Is Already Studying Medical AI in Real Healthcare Settings

Google has already begun related clinical work involving the text-based version of AMIE.

A feasibility study with Beth Israel Deaconess Medical Center has explored the technology's potential safety and usefulness in clinical practice.

Google is also involved in an ongoing nationwide randomized study with Included Health examining how AI could function within real-world virtual healthcare.

These studies could become much more important than controlled actor simulations because they will begin answering the central question surrounding medical AI:

Can these systems remain accurate, safe, reliable, and helpful when confronted with the complexity of real patients?

AI Doctors Are Getting Closer, But They Are Not Here Yet

AMIE represents another significant step toward multimodal medical AI capable of conducting something increasingly similar to an actual telemedicine consultation.

The technology can converse with patients, analyze visual information, listen for relevant signals, reason through possible diagnoses, recommend management options, and guide virtual examinations.

That is a major leap beyond the traditional symptom-checker chatbot.

But the distinction between clinical simulation and clinical medicine remains enormous.

Matching physicians in a controlled evaluation does not automatically mean an AI system is ready to diagnose millions of patients.

For now, Google's AMIE should be viewed as an advanced research system showing what the next generation of healthcare AI may eventually become.

And if future real-world trials confirm these early results, the question may eventually shift from:

"Can AI conduct a medical consultation?"

to:

"How should doctors, hospitals, governments, and patients use AI safely inside healthcare?"

That conversation is only beginning.