A real-world feasibility study at Beth Israel Deaconess found Google's diagnostic AI got the right answer into its shortlist 90% of the time, with zero safety interventions
Google’s medical AI just left the lab and met actual patients. The early report card looks surprisingly good.
In a prospective feasibility study at Beth Israel Deaconess Medical Center (BIDMC), Google’s AMIE system included the correct final diagnosis in its differential 90% of the time. Supervising physicians watched every conversation, and they never once had to step in.
That matters because most medical AI headlines come from benchmarks and simulations. This one came from 100 real adults with real symptoms, chatting with the AI before their real doctor’s appointments.
What happened at Beth Israel Deaconess #
AMIE, short for Articulate Medical Intelligence Explorer, is a research system built jointly by Google Research and Google DeepMind. It is designed for conversational diagnostic reasoning.
The study ran from April to November 2025. Each of the 100 adult participants talked with AMIE via text chat ahead of a primary care visit.
Live physician supervisors monitored those sessions against predefined safety criteria. The tally of required safety interventions came to zero.
Accuracy was measured after an 8-week chart review, which gave doctors time to confirm what patients actually had. The results broke down by how far down the AI’s list you had to look:
- **Top-7:** the correct diagnosis appeared in 90% of cases
- **Top-3:** 75% of cases
- **Top-1:** 56% of cases
The doctors and patients weighed in #
Primary care physicians received AMIE’s transcripts and summaries before seeing their patients. In 75% of cases, those PCPs reported feeling better prepared for the visit. Nearly 60% indicated the material could potentially change how they approached the patient’s care.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Patients responded very positively. Many said their attitudes toward AI improved after the interactions.
How AMIE got here #
AMIE was first introduced in 2024. Since then, it has been put through extensive testing in simulated settings.
Early studies found it met or exceeded board-certified PCP performance across a range of diagnostic dialogue metrics. According to specialist evaluators, AMIE came out ahead on 28 out of 32 axes covering diagnostic accuracy and consultation quality.
The system has kept expanding. Multimodal vision capabilities arrived by 2025, letting it work with visual inputs rather than text alone. Features for longitudinal disease management were reported in Nature in 2026. The research has also explored video consultations.
In a commentary published in Nature Medicine in September 2026, the developers argued that prospective real-world trials are essential to building trust in clinical AI. Benchmarks alone, they wrote, cannot establish efficacy and usability in real medical settings.
What this means for healthcare and the AI race #
The first caveat is the biggest one. AMIE remains an experimental research tool and has not been deployed in clinical settings.
One hundred patients at a single academic medical center is a feasibility study, not proof of safety at scale. Every conversation also had a physician watching in real time.
The 56% top-1 accuracy figure is also a useful reality check. It shows why the developers frame AMIE as a support tool and keep doctors firmly in the loop.
What to watch next: larger prospective trials across more sites and patient populations, studies without real-time physician supervision, and any regulatory pathway toward clinical deployment.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our