Google’s AMIE Tried Taking Patient Histories in a Real Clinic. Here’s What Happened.

Google Research and DeepMind have been working on AMIE (Articulate Medical Intelligence Explorer) for a while now. You might remember the hype around its diagnostic capabilities in simulated settings—talking to patient actors, helping clinicians with tricky cases. But simulations are safe, controlled, and frankly, not the real world.

Now they’ve taken the next step. In partnership with Beth Israel Deaconess Medical Center (BIDMC), they ran a prospective, single-center feasibility study where AMIE actually interacted with real patients in a real ambulatory primary care clinic. The paper is called “A prospective clinical feasibility study of a conversational diagnostic AI in an ambulatory primary care clinic,” and it’s a cautious but important milestone.

Here’s how it worked: Patients booked for new, non-emergency episodic complaints were invited to chat with AMIE before their actual doctor’s appointment. The AI conducted a text-based clinical history interview via a secure web link. But here’s the key safety net—a physician was watching the whole thing live via video call with screen sharing, ready to jump in if the AI hit any predefined safety tripwires. Think of it like a resident taking a history under an attending’s supervision.

The AI then generated a transcript and summary for the clinician to review before seeing the patient. This is the part that makes sense to me—offloading the tedious, repetitive history-taking that eats up so much of a doctor’s time. If AMIE can reliably gather the basics, that frees the physician to focus on the nuanced stuff: physical exam, diagnosis, shared decision-making.

But the real question is: Did it work? The study is described as a “feasibility” study, which is research-speak for “we’re checking if this thing crashes and burns before we bother measuring whether it helps.” The results are promising but not definitive. They reported high patient satisfaction scores, and clinicians found the summaries useful. But I’d take those numbers with a grain of salt—early adopter bias is real, and patients who volunteer for an AI study are probably more tech-friendly than the average person.

What I find more interesting is the safety data. The physician supervisors had to intervene in a small number of cases. The paper doesn’t go into gory detail about why, but any intervention is a flag. Was the AI asking irrelevant questions? Missing critical red flags? Being too pushy? Without specifics, it’s hard to judge how close AMIE is to being truly autonomous.

The other elephant in the room is the study design itself. This was single-arm, meaning no control group where patients just talked to a human. So we don’t know if AMIE actually saves time or improves outcomes compared to the current standard. That’s fine for a feasibility study—you walk before you run—but it means we’re still years away from this being a real product.

Google’s evidence roadmap here is sensible: start with simulations, then move to supervised real-world testing, then eventually randomized trials. But the gap between “feasible under close watch” and “safe enough to deploy at scale” is enormous. Clinical AI has a long history of looking great in controlled settings and falling apart when exposed to the messiness of real patients—different languages, health literacy levels, comorbidities, and emotional states.

AMIE handled the structured part of history-taking, but medicine is full of unstructured moments. A patient might mention chest pain only as an afterthought, or a symptom that sounds minor but is actually a red flag. Can AMIE pick up on those? The study doesn’t say, and I suspect that’s the next big hurdle.

Still, I’ll give Google credit for doing this the hard way. A lot of AI companies would have skipped the real-world testing and gone straight to a press release. AMIE’s approach—with physician oversight, IRB approval, and a pre-registered protocol—is how you build trust in a domain where mistakes can kill people.

The bottom line: AMIE can take a patient history in a real clinic without causing disasters, under close supervision. That’s not a slam dunk, but it’s a solid first step. The real test will be the next study—the one where they remove the training wheels and see if the AI can actually improve care without constant hand-holding.