A medical AI that can hold a live video conversation with a patient — listening, asking follow-ups, reasoning through a case — has matched primary care physicians in a controlled research setting.
Google's AMIE (Video) system earned clinical evaluator ratings on par with doctors across several core measures during simulated consultations. But the company is explicit: real-patient studies must follow before anyone draws conclusions about clinical use.
What Google's AMIE video consultation test actually showed
Fifteen trained actors portrayed patients across five broad condition groups: cardiopulmonary, abdominal, HEENT — head, eyes, ears, nose, throat — neurological or psychiatric, and musculoskeletal presentations. AMIE conducted synchronous video consultations with them, meaning the interaction happened in real time, not through delayed text exchanges.
Clinical evaluators then rated the AI's performance. On several core measures, AMIE's scores were comparable to those of primary care physicians handling the same consultations. Google has not indicated the AI outperformed doctors. The finding is parity — in a simulated setting.
Why an AI matching doctors on video matters for patients
The significance is not that an AI replaced a doctor in a test. It is that conversational, perceptive AI — capable of sustaining a natural video dialogue — has reached the threshold where clinical evaluation becomes meaningful.
If an AI can consult on par with a physician in simulation, the question of whether it can assist in real healthcare moves from theoretical to testable. That shift matters to anyone who has waited weeks for a primary care appointment or travelled long distances for a basic consultation.
Three agents, one consultation: inside AMIE's design
One of the most revealing details concerns how AMIE is built. Rather than assigning dialogue, clinical reasoning, and perception to a single model process, AMIE uses an asynchronous multi-agent architecture. A consultation is divided among three agents.
Google's reasoning is candid: a single agent cannot currently sustain natural conversation reliably. Splitting the work — one agent handling dialogue, another clinical reasoning, another perception — allows the system to manage the competing demands of a live video consultation. This architecture is AMIE's core technical differentiator, and it reflects a broader industry recognition that medical conversation requires specialised, coordinated AI processes rather than one general-purpose model.
How the actor-based study was built
The use of professional patient actors is a long-established method in medical education and clinical skills assessment. Actors are trained to portray conditions consistently, allowing evaluators to compare how different clinicians — or in this case, an AI — handle the same presentation under similar conditions.
The conditions spanned some of the most common reasons people seek medical care: chest discomfort, abdominal issues, ENT complaints, neurological or psychiatric symptoms, and musculoskeletal problems. The breadth suggests Google tested general-practice capability, not a narrow specialty.
Who could benefit if AMIE clears real-patient trials
The potential beneficiaries are people who struggle to access primary care — those in rural areas, on long waitlists, or with conditions that feel too serious to ignore but not urgent enough for an emergency room.
A system capable of conducting a competent initial video consultation could triage, reassure, or escalate, directing patients to the right level of care faster. None of that is available yet. But the direction of travel matters to anyone who has experienced the friction of getting a timely medical opinion.
What Google says about the road to clinical use
Google is careful not to overstate the findings. According to the research update, studies involving real patients and their own health conditions must be conducted before anyone can draw conclusions about AMIE's clinical use.
The simulated environment, however realistic, is not the same as a patient describing symptoms they are actually experiencing. The company has not announced a timeline for such studies in this update.
What this benchmark does — and doesn't — prove
The meaningful comparison is not AMIE versus a doctor. It is AMIE versus the previous generation of medical AI. Earlier systems could answer medical questions or generate notes after a visit. AMIE is being tested for the live, interactive part of medicine — the part that requires listening, observing, and responding in real time.
That is a categorically different challenge. The fact that Google needed three coordinated agents to meet it underscores how difficult natural medical conversation remains for AI.
Confirmed so far vs what remains unproven
Confirmed: AMIE conducted synchronous video consultations with 15 trained actors portraying conditions across five clinical areas. Clinical evaluators rated its performance on par with primary care physicians on several core measures. AMIE uses an asynchronous multi-agent architecture because a single agent cannot sustain natural conversation.
Not yet clear: whether AMIE's performance translates to real patients with actual conditions. How evaluators defined the "core measures" in detail. Whether parity holds in complex, emotionally charged, or emergency consultations. And any timeline for real-patient studies, which Google has not specified in this update.
The gap between trained actors and real patients
The limitations are substantial. Trained actors, however skilled, do not experience symptoms. They do not carry the anxiety of a patient hearing serious news. They do not present
