Google's AMIE Just Ran a Doctor's Video Visit. Here's What Actually Happened.
The research AI listened, looked, and diagnosed. Patient actors said they preferred it to a text chatbot.
Key points
- Google Research and Google DeepMind tested AMIE, a research medical AI, in real-time video consultations with simulated patients.
- AMIE is built on Google's Gemini model and Project Astra, using multiple specialised AI programs working together to see and reason.
- In a randomised study, clinical evaluators rated AMIE favourably against primary care physicians on history-taking, diagnosis, management and communication quality.
- Patient actors preferred the video experience over text-only chat.
- Google says AMIE is still research and not ready for real clinical use.
A doctor's visit is not just a conversation. Your GP watches how you walk in, hears the crack in your voice, notices you wincing when you sit down. That visual and audio layer is exactly what most medical chatbots miss.
Google wants to close that gap. This week the company showed off AMIE running a real-time video consultation, a first for a medical AI system of this kind. We first covered AMIE on 11 August 2026; this study is the most substantial test of it we've seen reported.
What did Google actually build?
AMIE talks to patients over video, watches for visual cues, and works through possible diagnoses as the conversation unfolds. It's not a product you can book an appointment with.
Under the bonnet, it runs on Gemini, Google's large language model, the same software that powers its consumer chatbot. It uses Project Astra, Google's system for handling live video and audio, and a multi-agent setup: several specialised AI programs working in parallel rather than one model doing everything alone.
In practice, that means AMIE can guide a patient through a virtual physical exam. Think: "Can you press there and tell me if it hurts?" or "Hold your hand up to the camera."
How did it perform against real doctors?
It held its own, at least in a lab setting. Google ran a randomised study with patient actors alongside a group of primary care physicians. Clinical evaluators then scored both sides.
AMIE was rated favourably on the things that matter in a GP visit:
| Clinical skill | What it means |
|---|---|
| History-taking | Asking the right questions about symptoms |
| Diagnostic accuracy | Working out what is likely wrong |
| Management | Recommending sensible next steps |
| Communication | Explaining things clearly and kindly |
Patient actors also said they preferred the video experience to text chat. That's a meaningful signal. A lot of telehealth today is still typing into a box.
Should patients get excited yet?
Not for a real appointment. Google is clear that AMIE is a research system and more work is needed before it can be responsibly used in a clinic. The study used actors, not sick people. Simulated cases are cleaner than the messy reality of a Tuesday morning surgery.
There's also survivorship bias to watch for in AI medical demos. We see the impressive runs; the transcripts where the model missed something a nurse would spot in ten seconds don't make the press release.
Still, this is a real step. Until now, most medical AI research has focused on text: symptom checkers, note summarisers, chat interfaces. Adding live vision and voice gets closer to how doctors actually work.
What does this mean for ordinary people?
In the short term: nothing changes. Your next GP appointment will still be with a human.
In the medium term, expect telehealth services to start experimenting with video-aware AI assistants, probably first as a helper for the doctor rather than a replacement. That could mean shorter waits and better triage before your GP even joins the call.
Honest judgement on this one: the lab results are genuinely impressive, but the gap between actor-played scenarios and a real patient who can't describe their symptoms well is enormous. Watch whether Google moves toward real-world trials, and watch who carries the liability when it gets something wrong.
