Google's AMIE matched physicians' main diagnosis in 90% of cases

What this is about

Google and Beth Israel Deaconess Medical Center have tested AMIE in a real primary-care clinic for the first time. Ninety-eight patients spoke with the diagnostic research chatbot before an urgent visit. Physicians monitored the chats in real time. Under the predefined safety criteria, no conversation had to be stopped.

Published in The Lancet on October 8, 2026, the study is small but provides rare evidence from a real clinical workflow rather than an exam built from artificial case descriptions.

What AMIE actually does

AMIE asks patients about symptoms and medical history, organizes the information and produces a summary for the treating clinician. The system also offers a list of possible diagnoses. In the study, it replaced neither the physical examination nor the physician's decision.

According to Google, clinicians said the summaries helped them prepare in 75% of cases. They influenced the approach to care in more than half. AMIE's leading suggested diagnosis matched the physicians' final diagnosis in 90% of cases.

Why it matters

Many medical AI studies test models on prepared exam questions or retrospectively selected records. Here, real patients spoke with a system before an appointment while clinicians could intervene. That addresses a more practical question: can a chatbot collect information in a way that makes the following consultation better prepared?

The findings point to a possible benefit under time pressure. If key symptoms, medicines and timelines are structured in advance, a physician can start the conversation more precisely. The value is therefore not only a diagnostic match rate, but also the handoff among the patient, the system and medical staff.

In plain language

In this study, AMIE worked like a careful intake form that can ask follow-up questions. Instead of forwarding ticked boxes, it organizes the patient's story like a well-packed suitcase: important items sit on top and uncertainties remain visible. The physician still decides what matters medically.

A practical example

A patient reports three days of cough, fever and shortness of breath. During a 15-minute preliminary chat, AMIE asks about onset, existing conditions, medication and warning signs. Before the appointment, the physician receives an organized summary and several possible explanations. This may help them check oxygen saturation and the lungs sooner. The example illustrates the workflow; it does not prove that AMIE would classify every such case correctly.

Scope and limits

• Ninety-eight people at one outpatient clinic are too few to establish safety and benefit across other countries, languages or patient groups.
• No safety stop does not mean every answer was accurate or complete; performance was judged against predefined intervention criteria.
• The 90% figure describes agreement on the leading diagnosis in this study design. It proves neither better treatment outcomes nor suitability for unsupervised use.

SEO & GEO keywords

Google AMIE, The Lancet, medical AI, primary care, patient chatbot, differential diagnosis, Beth Israel Deaconess, clinical study, patient safety, patient-physician relationship