Google’s medical AI can now run a video consultation. The study was with actors.
AMIE reads visual and audio cues and guides a virtual physical examination. Clinicians rated it well — against simulated patients, in a research setting Google says is not ready for real use.
Google has published research showing AMIE, its experimental diagnostic AI, conducting consultations over video — interpreting what it sees and hears, guiding a patient through a virtual physical examination, and reasoning about a diagnosis in real time.
In a randomised study using simulated consultations, clinical evaluators rated the system favourably on history-taking thoroughness, diagnostic accuracy, management appropriateness and communication quality. Patient actors preferred the video experience to text chat.
The last sentence of Google’s own post is the one that governs how the rest should be read: “AMIE remains a research system and more research is needed before responsible real-world clinical deployment.”
What “simulated consultations with patient actors” means
The patients were actors working from scripted scenarios. This is a standard and respectable method — it is how medical schools examine students, because it produces a controlled comparison that real patients cannot.
It also has known limits. An actor performs a condition they have been briefed on. They do not have twenty years of confused history, three other conditions, a medication list they cannot recall, or the specific quality of not-quite-answering that real patients bring. They are not frightened. They do not minimise symptoms because they are worried what the answer will be.
Diagnostic reasoning is hardest exactly where those things live. A strong performance against scripted presentations is a real result and a necessary step — it is not evidence of performance in a clinic, and Google does not claim it is.
The comparison that would settle it
Google’s post reports that evaluators assessed AMIE favourably across core competencies. What it does not put front and centre is a head-to-head figure against the primary care physicians in the study — how often each reached the correct diagnosis, on the same cases, scored the same way.
Earlier AMIE work, published in Nature, did make that comparison directly in a text-based setting. For the video results, the detailed breakdown is what to look for when the full paper appears. The gap between “rated favourably by evaluators” and “outperformed clinicians on N cases” is the entire question, and the first phrasing is compatible with a wide range of second ones.
What would have to happen before it saw a patient
In Britain, software that diagnoses is a medical device. It falls under the MHRA’s framework for software and AI as a medical device, and deployment inside the NHS would additionally require compliance with clinical safety standards governing how such systems are introduced and monitored.
That is a multi-year path involving evidence from real clinical settings, not simulated ones. Nothing in this announcement shortens it, and Google is not suggesting otherwise.
Why it is still worth attention
The interesting claim is not that an AI diagnosed well. It is that it handled a video consultation — reading visual signs, prompting someone to move a limb or point a camera, and adjusting as it went.
Remote consultation is now a permanent part of how medicine is delivered, and its weakness has always been the missing physical examination. Anything credible in that gap matters, whoever eventually builds it. That is worth watching closely and reporting carefully, which is not the same as reporting it as a doctor.
It is also not the only route by which consumer technology is arriving in clinical territory. Samsung has just had a hearing aid feature on its earbuds cleared by the US regulator — a far smaller claim than diagnosis, and correspondingly much closer to shipping.