An AI doctor can now conduct a live video consultation while listening to speech, watching the patient and reasoning about what it sees in real time. In a new preprint released on 10 August 2026, Google’s AMIE research system was compared with 30 primary care physicians across 100 simulated clinical scenarios. Specialist evaluators rated the video version of AMIE on par with or better than doctors across history-taking, diagnosis, management and visual observation, while patient actors preferred how the AI assessed and explained conditions. They still preferred human doctors for rapport and partnership.
That split is the most important result. The real-time video consultation study does not show that an AI doctor is ready to replace a GP, and it did not test autonomous care on real patients. It shows something narrower and potentially more consequential: medical AI is moving beyond typed symptom checkers into the sensory territory of an ordinary telehealth appointment, where tone of voice, movement, appearance and what a patient can physically show the camera become part of the clinical conversation.
How the AI Doctor Video Consultation Actually Worked
AMIE stands for Articulate Medical Intelligence Explorer. The new video configuration is a Gemini-based multi-agent system designed to combine low-latency conversation, clinical reasoning and real-time audio-visual perception. Instead of waiting for a patient to describe everything in text, the system can use what it hears and sees during the call while deciding what to ask next.
The researchers built a taxonomy of clinical audio-visual cues and then evaluated AMIE Video in a randomised Objective Structured Clinical Examination, or OSCE. OSCEs are widely used in medical training to test clinicians against standardised scenarios. This study involved 30 primary care physicians, 15 trained patient actors and 100 scenarios. The actors consulted AMIE Video, a text-only version of AMIE and human doctors under controlled conditions.
Clinical evaluators rated AMIE Video as comparable to or better than the physicians on the study’s measures of history-taking, diagnosis, management and physical observation or examination. Patient actors preferred AMIE’s approach to assessing and explaining conditions, and they preferred the video interface over text chat for communicative effectiveness, convenience and feeling understood. Yet human doctors retained the advantage in rapport and partnership-building. That is a more revealing result than a simple league table of AI versus doctors.
What This Means for You
If you are a patient, this research does not mean you should replace a medical appointment with Gemini or another general-purpose chatbot. AMIE is a research system built specifically for clinical dialogue, the August study used simulated cases rather than ordinary patients, and the authors explicitly say further research is needed before real-world translation. The safe takeaway is about direction of travel, not present-day substitution.
The direction is clear. A future telehealth system may be able to ask you to turn your head, show a rash, demonstrate how a joint moves, describe a pain out loud and answer follow-up questions while continuously combining those signals into its reasoning. That is qualitatively different from typing symptoms into a chat box. LiveAIWire’s earlier examination of the AI doctor dilemma argued that implementation and oversight determine whether diagnostic AI becomes useful support or a patient-safety hazard. Real-time video makes that governance question more urgent because the system can collect and interpret a much richer stream of information.
For now, the sensible patient question remains whether a qualified clinician is responsible for the final decision. A system that helps a doctor gather information, structure a history or notice a visual cue is a different proposition from software independently diagnosing and managing an illness. The new AMIE results are evidence about capability. They are not regulatory approval, proof of safety at population scale or evidence that autonomous AI care produces better health outcomes.
The AI Can Now See Things You Struggle to Describe
Text is a poor substitute for many parts of medicine. A patient can say that a rash is red, but a clinician may also care about its shape, distribution, borders and whether it changes when pressed. Someone can describe a tremor, but video can show its rhythm and amplitude. Speech can convey breathlessness, hesitation or changes in voice that are awkward to translate into a typed message.
AMIE’s progression towards video did not happen in one jump. In earlier work published in Nature Medicine in 2026, a multimodal version of the system was tested on 105 simulated telehealth scenarios involving uploaded material such as skin photographs, electrocardiograms and clinical documents. Eighteen specialist physicians evaluated the consultations. Multimodal AMIE performed similarly to or better than primary care physicians across the assessed comparison axes and showed higher diagnostic accuracy in that study.
The video research extends that idea from uploaded artefacts to a continuous interaction. Instead of examining a still image after it arrives, the system has to decide what to look for while the consultation is happening. That introduces a harder problem: perception and conversation have to work together quickly enough that the exchange still feels natural.
But Seeing a Patient Is Not the Same as Examining One
The paper’s limitations matter. The researchers report weaknesses in fine anatomical precision, subtle affective cues and high-frequency movements. A camera also cannot reproduce palpation, reliably measure many vital signs, listen through a stethoscope or replace tests that require equipment and trained handling. Even human telemedicine has limits for exactly these reasons.
There is another limitation hidden inside the phrase “video consultation”. What an AI sees depends on the camera, lighting, connection quality, angle and what the patient knows how to show. A beautifully controlled evaluation can demonstrate what the system is capable of under test conditions without proving that the same performance will survive a dim bedroom, an unstable mobile connection or a patient who cannot position the camera correctly.
This is why LiveAIWire’s broader review of AI-powered healthcare keeps returning to the difference between a strong model and a strong clinical deployment. Medicine is not only an inference problem. It is a workflow involving data quality, human judgement, accountability, follow-up and the ability to recognise when the available information is inadequate.
The Most Surprising Result Was Not Diagnostic Accuracy
The headline-grabbing finding is that specialist evaluators rated the AI very highly. The psychologically more interesting result is the division in patient-actor preference. AMIE was preferred for aspects of assessing and explaining the condition, but doctors were preferred for rapport and partnership. In other words, technical conversational competence and human relationship-building separated rather than moving together.
That distinction could shape how medical AI is deployed. A system that is exceptionally patient, systematic and clear at gathering a history may be valuable precisely because a human clinician does not have to surrender the relational part of medicine. The strongest future model may therefore be neither “AI doctor” nor “human doctor”, but a consultation in which AI quietly handles parts of information gathering and reasoning while a clinician retains responsibility and the human relationship.
There is already some real-world evidence for that more limited role. In a separate prospective feasibility study published in March 2026, 100 adults used a text version of AMIE before urgent-care appointments at Beth Israel Deaconess Medical Center. Human safety supervisors monitored every AI interaction and did not need to stop any consultation under the study’s predefined criteria. Patients reported high satisfaction, and primary care physicians found the output useful for preparation.
That real-patient study also supplied a useful reality check. AMIE’s differential diagnosis included the final chart-reviewed diagnosis in 90 percent of cases and placed it in the top three in 75 percent. Yet doctors produced management plans rated better for practicality and cost effectiveness. The AI could reason impressively while still missing parts of clinical judgement that matter when recommendations have to work in an actual healthcare system.
An AI Doctor Has to Manage Time, Not Just Symptoms
Diagnosis is only one part of primary care. Patients return, treatments fail, symptoms evolve and new test results change the picture. In June 2026, a peer-reviewed Nature study of AMIE for longitudinal disease management compared the system with 21 primary care physicians across 100 simulated scenarios, each involving three visits. AMIE’s management reasoning was non-inferior overall and scored better on several measures of treatment precision, investigations and guideline alignment.
The authors were equally clear about what the result did not establish. Those encounters were simulated and conducted through text chat. They concluded that further prospective work with real patients and appropriate safety oversight would be required before clinical use. The video study therefore sits within an evidence ladder: text simulations, multimodal simulations, a supervised real-patient feasibility study, longitudinal reasoning tests and now real-time audio-visual simulation.
That sequence is encouraging because it is the opposite of releasing a consumer chatbot first and discovering the risks later. It also means the evidence should be read stage by stage. Success in an OSCE supports further evaluation. It does not automatically support deployment.
Why Patient Preference Needs Careful Reading
“Patients preferred AI” would be a seductive conclusion and a misleading one. The participants making the comparison in the August video study were trained patient actors following clinical scenarios, not people seeking care for their own illnesses. Their preferences tell us something about consultation quality under controlled conditions, but they cannot establish how frightened, elderly, seriously ill or medically complex patients would respond to the same system.
The preference was also not universal. The actors favoured AMIE in specific areas, particularly assessment and explanation, while favouring physicians for rapport and partnership. That suggests different components of a consultation can be judged separately. An AI may explain a differential diagnosis very clearly while still failing to create the sense that another person understands the stakes of being ill.
That is particularly relevant to fields where trust changes whether patients follow advice. LiveAIWire’s analysis of AI cancer diagnosis found that raw accuracy is only part of adoption. Patients and clinicians also need to understand who reviews an algorithmic finding, how bias is monitored and who is accountable when the system is wrong. A video interface makes AI feel more like a clinician, but a more human interface does not create human accountability.
The Privacy Stakes Rise When the AI Can Watch You
A text medical conversation is already sensitive. Real-time audio and video add voice, appearance, movement, surroundings and potentially other people in the room. Any eventual clinical deployment would therefore have to answer not only whether the system is accurate, but what is recorded, what is retained, who can access it and which parts of the audio-visual stream are used for model improvement.
The August paper establishes a research capability, not a consumer privacy regime. It would be premature to infer future product policies from an experimental system. But the direction of travel changes the privacy question from “what did you tell the AI?” to “what could the AI observe while you were talking to it?” That is a materially larger information surface.
Healthcare systems already handle extremely sensitive records under strict legal and professional obligations. Adding continuous machine perception to telemedicine will require equally explicit rules around consent, retention, auditability and human review. Those governance questions are not secondary to clinical accuracy. They are part of whether the technology is safe enough to use.
The Doctor Is Not Disappearing. The Consultation Is Changing
The most credible interpretation of AMIE’s video results is not that the GP has been beaten. It is that several abilities once bundled together inside one human consultation can now be technically separated. History-taking, visual observation, explanation, diagnostic reasoning, management planning, rapport and accountability do not have to be performed by the same actor in the same way.
That creates opportunities and risks. An AI that gathers a more complete history before a clinician joins could give doctors more time for examination, difficult decisions and conversation. An AI that independently conducts the whole encounter could instead remove precisely the human partnership that patient actors still preferred. The technical capability does not decide which model healthcare systems choose.
LiveAIWire’s coverage of AI physiotherapy has shown the same pattern in another part of care: the evidence is strongest when AI extends professional capacity and weaker when technology is marketed as a standalone substitute for professional judgement. AMIE’s latest result should be judged by the same standard.
The AI doctor can now see and hear enough to participate in something that looks far more like a medical consultation than a chatbot session. That is a genuine technical milestone. The harder milestone will be proving, with real patients in real health systems, that adding those senses improves care without weakening safety, privacy, accountability or the human relationship that the experiment itself suggests people still value.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.