Google's medical chatbot interviewed 98 real urgent-care patients, and doctors never had to step in

Google and Beth Israel Deaconess Medical Center published a study in The Lancet on Thursday, Oct. 8, 2026, testing AMIE, Google's research diagnostic chatbot, with real patients. Ninety-eight adults chatted with AMIE before urgent primary-care visits while physicians watched every conversation live. None of the chats had to be stopped for safety, and most doctors who read AMIE's notes said they helped them prepare.

Medical AI has spent years acing exams and simulated patients, which is a bit like judging a driver on a video game. This is the first time Google's diagnostic chatbot sat in a real clinic's front door, talking to real people with real sore throats and back pain, and the headline result is pleasantly boring: no conversation needed a doctor to yank the plug. That boring part is the point. The fair limits are big ones: it's 98 patients at one Boston clinic, every chat had a physician babysitter watching live, there was no control group, and only 44 doctors read the transcript before walking in. So this doesn't show a chatbot can replace triage. It shows something more practical and more likely to arrive first: an AI that does the 10-minute intake interview before you get to the exam room, so your doctor walks in already knowing why you came.

Google and Beth Israel Deaconess Medical Center (BIDMC) in Boston published the study in The Lancet on Thursday, 8 October 2026. Google says it is its first publication in the main Lancet journal. The paper is titled “Conversational diagnostic artificial intelligence in ambulatory primary care: a prospective feasibility study.” Those lines are Google’s and The Lancet’s.

Design. It was a prospective, single-center, single-arm feasibility study at BIDMC’s academic primary-care practice. Prospective means they watched new patients going forward, not old charts. Single-center means one clinic. Single-arm means there was no comparison group that skipped the chatbot. Feasibility means a first test of whether it can be done safely. English-speaking adults chatted by text with AMIE — Articulate Medical Intelligence Explorer — up to five days before an urgent primary-care appointment for a single complaint. Physician safety supervisors watched every conversation in real time and were trained to stop it on predefined safety criteria. AMIE’s transcript and summary were shared with the patient’s doctor before the visit. Those lines are the authors’.

From April to November 2025, 114 patients enrolled and 98 completed both the AMIE chat and the doctor visit. Those counts are the authors’.

Zero conversations had to be stopped for safety. Supervisors noted one hallucination — a made-up clinical detail — and added clinical information in five conversations. Those lines are the authors’.

Clinical evaluators rated AMIE’s conversations favorably in 87% to 100% of cases across 17 criteria. Patients rated them favorably in 48% to 96% across 16 criteria. Patients’ attitudes toward AI improved after using AMIE and stayed higher after their doctor visit. Those scores are the authors’.

Doctors completed follow-up surveys in 60 of 98 cases, including 44 where they had read AMIE’s transcript beforehand. In those 44, they found it helpful for preparing in 33 cases (75%) and said it might have changed how they handled the visit in 25. Those figures are the authors’.

AMIE’s list of possible diagnoses included the final diagnosis, from a chart review after the visit, 90% of the time. A list of possible diagnoses is the set of conditions the chatbot thought might explain the complaint. Those lines are the authors’.

AMIE is a research system, not a product. The authors and Google both say larger trials are needed before patient-facing AI like this is used at scale.

In plain terms, Google and a Boston teaching hospital published a Lancet study in which 98 adults texted with AMIE before an urgent primary-care visit, with a doctor watching every chat. None of the conversations had to be stopped. In the 44 visits where the doctor had read the transcript, 33 said it helped them prepare. The chatbot’s list of possible diagnoses included the final one 90% of the time. It is not a product, and there was no control group.