A 98-patient AMIE study at one clinic reports no safety stops before urgent visits

On 8 October 2026, Mike Schaekermann wrote on the Google blog about a clinic study of AMIE, the Articulate Medical Intelligence Explorer. Google’s research page calls AMIE a research project on clinical reasoning and conversations, shared across Google Research, Google DeepMind, and other Google health groups. The page does not offer it as a product.

The paper is “Conversational diagnostic artificial intelligence in ambulatory primary care: a prospective feasibility study,” DOI 10.1016/S0140-6736(26)01535-7. The PubMed abstract (PMID 42849491) is dated 8 October 2026 in The Lancet. Peter G. Brodeur and colleagues affiliate with Beth Israel Deaconess Medical Center, Google Research, and Google DeepMind. The funder line is Alphabet. The blog links the Lancet page.

One clinic, before the visit

The abstract calls the work a prospective, single-centre, single-arm feasibility study. English-speaking adults talked with AMIE up to five days before a single-complaint urgent visit in BIDMC ambulatory primary care. Physician safety supervisors watched and were trained to interrupt on predefined safety criteria. Transcripts and summaries went to the primary care physician before the visit. There is no comparison arm.

A differential diagnosis is the list of conditions a clinician keeps open as possible explanations for a complaint until more information narrows it. The study reports how those pre-visit chats and summaries lined up with the visit. It does not put AMIE in the role of the physician who makes the diagnosis.

From April to November 2025, the abstract says, 114 patients enrolled and 98 completed both the AMIE conversation and the appointment. Primary outcomes were supervised safety stops, conversation quality scored by clinical evaluators, and surveys of patients and physicians. The registry is NCT06911398. That record calls the study a completed observational prospective cohort at BIDMC, started 2 April 2025, with actual enrollment of 100. The registry’s 100 and the abstract’s 114 enrolled and 98 completed are not the same count.

The figures, as each page states them

The abstract says: “Zero conversation safety stops were required on the basis of prespecified criteria.” Schaekermann writes: “not a single conversation needed to be interrupted based on predefined safety criteria.” The abstract adds that supervisors noted one hallucination and added clinical information in five interactions. Zero stops under this protocol is a count from this study.

Evaluators rated conversations favourably in 87 to 100 percent of cases on 17 criteria, and patients in 48 to 96 percent on 16 criteria. Physicians returned surveys in 60 of 98 cases, including 44 who reviewed the transcript first. The abstract says: “PCPs found AMIE helpful for visit preparation in 33 of 44 cases and reported that it might have changed their behaviour in 25 of 44 cases.”

The blog compresses that to summaries that “helped them prepare for visits in 75% of cases” and “influenced their approach to care in more than half.” Thirty-three of 44 is 75 percent. Twenty-five of 44 is more than half, and the abstract’s verb is “might have changed.” The blog also says AMIE’s differential diagnoses “matched the doctors’ final diagnoses 90% of the time.” That 90 percent is not in the PubMed abstract.

The abstract says further research is needed and calls the result initial feasibility in one real setting. Schaekermann writes that larger clinical trials are needed before anyone assesses patient-facing AI at scale. The counts belong to this supervised study: one clinic, urgent primary care, and a preparation question answered by 44 physicians.