Google's AMIE Meets Real Patients: What a Lancet Feasibility Study of 98 Primary Care Patients Does and Doesn't Show

This article is general information about a research study, not medical advice. Talk to a clinician about your own health.

Google published a study in The Lancet on Oct. 8, 2026 testing its conversational diagnostic AI, AMIE, with real patients in primary care. Titled "Conversational diagnostic artificial intelligence in ambulatory primary care: a prospective feasibility study," it is described by Google as its first main-edition Lancet publication. We read Google's announcement and secondary analyses, not the paper itself, so the details below are as those sources reported them.

What the Study Did

• Design: A prospective, single-centre, single-arm study at Beth Israel Deaconess Medical Center, run from April to November 2025. Single-arm means there was no comparison group.
• Participants: 114 patients enrolled, and 98 completed both the AI interaction and the appointment.
• The interaction: Patients chatted with AMIE by text up to five days before their visit. The system gathered history and produced information for the clinician.
• Safety supervision: A physician safety supervisor watched each session live by video.

What It Found

• Safety: No conversation was stopped on any of four pre-specified safety criteria. One hallucination was noted, and clinical information was added to the record in five interactions.
• Usefulness to clinicians: Clinicians found the preparation helpful in 33 of 44 relevant responses, or 75%. Google says it influenced the clinician's approach in more than half of cases.
• Diagnostic accuracy: Google's post describes AMIE's differential diagnoses as matching the doctors' final diagnoses about 90% of the time. Secondary analysis says the 90% figure refers to the correct diagnosis appearing within AMIE's top seven suggestions, as reported in a March 2026 preprint, and that the top-three figure was about 75%. We have not confirmed the exact framing in the paper.

What It Does and Doesn't Show

This section is analysis, not reporting.

• It shows feasibility under supervision. A physician watched every session, so the study says little about how the system behaves alone.
• There is no comparison. Without a control group, such as visits without AMIE, the study cannot show that patients did better or that clinicians were more accurate or faster.
• The sample is selective. Only 98 of 114 enrolled patients completed the process, and completers may differ from those who dropped out.
• The funder is the developer. The study was funded by Alphabet, Google's parent, which is common for early studies but means independent replication matters.
• Google says larger trials are needed. The AMIE team's own commentary in Nature Medicine, as quoted by AI Weekly, says: "Trust in clinical artificial intelligence (AI) cannot be benchmarked into existence."

That last point is the most useful framing. Benchmark scores and preprints told us AMIE could perform on written cases. This study is the first step from there to real patients, and the next steps would be controlled trials that measure outcomes, not impressions.

Why It Matters

• Medical AI is entering real clinics. The test used real patients and real appointments, with oversight.
• Safety supervision is part of the design. A live physician watching each session is a cost that would not scale as-is.
• The key evidence is still missing. No comparison group means no evidence yet on outcomes.
• Wording matters. "90% match" and "top seven suggestions" are different claims, so check what a figure measures.