AI news · October 12, 2026

Google studies AMIE with 98 completed primary-care patients

healthcareresearchagentsevaluation

Google's AMIE study had 98 completed patients, one reported hallucination and no safety stops; builders should copy the supervised pre-visit summary workflow, not sell autonomous diagnosis.

patients who completed the study
98
reported hallucination
1
clinical evaluator favourable ratings
87–100%
Made With Models illustration for this story

Google reports a prospective, single-centre feasibility study in which AMIE interviewed adults for up to five days before an urgent primary-care appointment. Of 114 people enrolled, 98 completed the study; physician safety supervisors monitored the conversations, and there were zero safety stops. The paper reports one hallucination, clinician-added information in five cases, and favourable ratings from clinical evaluators in 87–100% of assessed conversations. Primary-care doctors found the system helpful in 33 of 44 encounters and said it could change their approach in 25.

This is a narrow supervised workflow, not evidence that an agent can diagnose safely on its own. Google also reports that AMIE's differential diagnosis matched the final diagnosis in 90% of cases, but the sample is small and the setting is controlled. The useful comparison is a structured intake and summary tool, not a replacement for a clinician. The next proof point is replication across clinics, languages, conditions and real operating constraints.

What you can do with it

Read the study and prototype the low-risk part: collect a pre-visit history, show the patient a draft summary, and give the clinician the source conversation plus uncertainty flags. Keep a human reviewer in the loop and measure missing facts, correction time and unsafe suggestions before adding any decision support.

Our take

The important product lesson is workflow placement. A model that gathers context before an appointment can save time even when a clinician remains responsible. The headline accuracy figure is less decisive than the one hallucination and the small, monitored sample; healthcare builders should sell preparation and documentation first.

Source: Google Research / The Neuron — Made With Models writes the brief; the reporting is theirs.