Google's AMIE, a research chatbot built to take medical histories and reason about diagnoses, held pre-visit conversations with 98 real patients at a Boston primary care clinic without a single conversation being stopped for safety reasons, the company said Thursday. The results were published in The Lancet, and Google described them in a post on its The Keyword blog.

What Happened

The study took place at the ambulatory primary care clinic of Beth Israel Deaconess Medical Center. Patients who were booked for urgent care appointments chatted with AMIE, short for Articulate Medical Intelligence Explorer, before seeing a doctor. Supervising physicians followed each conversation in real time and were ready to intervene.

According to Google, none of the 98 conversations met the study's predefined criteria for interruption. AMIE then produced summaries for the clinicians who saw the patients. Those clinicians reported that the summaries helped them prepare in 75% of cases and changed how they approached care in more than half of the visits.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

The system's differential diagnosis, the ranked list of conditions it considered possible, matched the doctors' final diagnoses 90% of the time, Google said. The post was written by Mike Schaekermann, a research lead at Google, who noted that this is the company's first paper in the flagship Lancet journal rather than one of its specialty titles.

Why It Matters

Most earlier AMIE research tested the system in simulated consultations. In studies Google began publishing in 2024, trained actors played patients and typed with either AMIE or primary care physicians in a format modeled on the structured exams used to assess medical trainees. Those results drew attention because AMIE often scored well against doctors, but they said little about how the tool behaves with real people who have real symptoms and incomplete stories.

The Beth Israel study is a narrower test, but a more realistic one. It asks whether a conversational AI can safely sit at the front of a clinical visit, gathering history before the patient reaches a physician. That intake role is where health systems see the clearest near-term use for patient-facing chatbots, because it could save clinician time without handing the AI any final decision.

The design also matters. Every conversation had a physician watching live, so the study shows what happens under close supervision, not what would happen if patients used the tool alone. The safety result is a count of conversations that did not trigger a stop, not proof that the AI never erred.

Limits of the Evidence

Google itself frames the work as an early step. The company said larger clinical trials are needed to evaluate patient-facing AI at scale. The blog post does not report patient satisfaction scores or a comparison group of visits without AMIE, and 98 patients from one clinic in one city cannot show how the system handles a broader mix of ages, languages, conditions and care settings.

The 90% figure also measures agreement with the treating doctors' conclusions, which depended on the full in-person visit. It does not show how often AMIE would reach the right answer on its own, or how its suggestions might steer physicians who read the summary first.

What to Watch

The next milestone is a larger trial, which Google says is required before conclusions can be drawn at scale. Key questions include how AMIE performs with less direct oversight, whether its summaries save measurable clinician time, how patients judge the experience and how regulators classify a tool that talks to patients before a doctor does. Google has not said when or where a follow-up trial would run.