Google-BIDMC AMIE Study Published in The Lancet

Google has published its first research paper in the main journal of The Lancet, describing a real world test of AMIE, the company’s research diagnostic chatbot, inside an active primary care clinic. The study was carried out with physicians at Beth Israel Deaconess Medical Center in Boston and tracked 98 patients who spoke with AMIE before urgent ambulatory care visits, rather than in a simulated or lab setting.

Supervising physicians watched every conversation as it happened, and the paper reports that none needed to be stopped under the safety rules set out before the trial began. That detail matters because most prior AMIE research, including earlier published work, relied on actors or recorded consultations rather than people with a genuine, same day medical need. Letting the tool speak directly with real patients ahead of a live visit, with a doctor able to intervene, is a different bar from a controlled demonstration.

Clinicians involved in the study also reported practical value from the AI generated summaries that AMIE produced after each conversation. In three quarters of cases, doctors said the summary helped them prepare for the visit, and in over half it changed how they approached the patient’s care. Separately, the differential diagnoses AMIE proposed lined up with the physician’s eventual diagnosis in nine of every ten cases, a figure Google is presenting as evidence the system can reason about a patient’s symptoms with real clinical accuracy rather than just producing plausible sounding text.

Google is careful to frame this as an early signal rather than proof the approach scales, noting that larger trials are still needed before any claim about broader deployment follows. There is no mention of a product timeline, a cost estimate or a plan to offer AMIE directly to patients outside research settings.

The numbers

• 98 patients consulted AMIE ahead of urgent primary care visits, each conversation supervised live by a physician.
• Clinicians said AI generated summaries helped them prepare for the visit in 75% of cases.
• AMIE’s differential diagnosis matched the physician’s final diagnosis in 90% of cases reviewed.

For health systems and insurers watching this space, the headline is not that an AI chatbot can produce a plausible diagnosis. It is that a hospital was willing to let one speak with real patients unscripted, under live physician supervision, and report the result in a leading medical journal rather than a company blog post alone.