Google has tested a research version of its medical AI system, AMIE, in live video consultations with professional patient actors. Independent clinical evaluators rated the system on par with primary care physicians across several measures, including history-taking, diagnostic accuracy, management recommendations and communication.
The findings are notable because the system had to speak naturally while interpreting audio and video and updating its clinical reasoning. They are not evidence that AMIE can safely treat real patients. Google explicitly says studies involving people with their own health conditions must come next before conclusions about clinical deployment are justified.
Three agents divide the demands of a consultation
AMIE uses separate “talker,” “planner” and “perception” agents rather than asking one model process to handle every task. Google says a single agent cannot currently maintain natural response times while also performing detailed reasoning and continuously processing audiovisual information.
The talker conducts the spoken exchange and uses information supplied by the other components. In the background, the planner updates possible diagnoses and management options, identifies missing information and reprioritises clinical objectives. The perception agent monitors the video and audio streams for non-verbal cues, physical findings and sounds that may matter clinically.
This division is partly a response to latency. Deep reasoning can create pauses long enough to damage the rhythm and rapport of a medical conversation. Separating the patient-facing dialogue from slower analysis allows AMIE to answer while other processes continue working.
Google reports that automated evaluations linked each agent to improvements in clinical measures, communication and response time. That result supports the architecture within the company’s tests, but it does not establish how reliably the components would coordinate across unpredictable real consultations or technical interruptions.
Controlled comparisons produced encouraging scores
The human evaluation included 15 trained actors portraying cardiopulmonary, abdominal, neurological or psychiatric, musculoskeletal, and head, eye, ear, nose and throat presentations. AMIE completed synchronous video consultations, while a text-only AMIE version provided a comparison. Ten board-certified primary care physicians used the same video interface.
Twenty experienced primary care doctors independently reviewed the consultations using general and case-specific clinical criteria. According to Google, evaluators rated video AMIE similarly to the physician group for thorough histories, diagnostic accuracy, appropriate management and communication. It matched or exceeded the text system on the same measures.
The video version received higher scores for identifying physical signs and proactively guiding actors through virtual examination manoeuvres than either physicians or text-only AMIE. Patient actors also preferred video to text and rated the system favourably for ease of use, rapport, empathy and confidence in care.
Those comparisons should be interpreted within the study design. The actors followed prepared scenarios, and evaluators scored recorded encounters rather than patient outcomes. A controlled examination can reveal whether a system performs defined behaviours consistently; it cannot reproduce the uncertainty, incomplete histories or emotional complexity of everyday medicine.
Real patients remain the decisive test
Before the physician-reviewed study, Google used an automated evaluation suite based on telehealth competencies from medical literature. Single-turn tests examined tasks such as recognising respiratory distress or anatomical laterality. Longer simulations assessed conversation, sometimes representing visual input through text descriptions rather than a complete live video feed.
That process helped expose capability gaps and refine the system quickly. It also illustrates the distance between development tests and clinical reality. Google reports occasional perception and reasoning errors, intermittent disruptions to conversational naturalness and broader technical questions around the prototype infrastructure supporting AMIE.
Professional actors cannot represent every symptom authentically, especially when audiovisual perception carries substantial diagnostic weight. Real patients also bring varied connectivity, surroundings, medical histories and behaviour that prepared cases do not capture.
Google has begun related clinical work with the text-based AMIE, including a feasibility study with Beth Israel Deaconess Medical Center and an ongoing nationwide randomised study with Included Health. Those projects concern a related modality, not proof for the video system described here.
The experiment therefore demonstrates a promising interaction design: fast conversation supported by slower specialist processes. Whether that architecture becomes a safe clinical tool will depend on evidence from real patients, measured outcomes and governance capable of handling errors—not evaluator parity in a simulated consultation alone.
