Google and BIDMC report a supervised pre-visit AI chatbot study in The Lancet

Google and Beth Israel Deaconess Medical Center (BIDMC) published a study in The Lancet on October 8, 2026, testing whether a research chatbot can talk with patients before they see a doctor. Google's research lead for the work, Mike Schaekermann, wrote in a company blog post that it is Google's first publication in the main Lancet journal.

The setting was narrow. At BIDMC's ambulatory primary care clinic, 98 patients consulted AMIE, which Google calls its research diagnostic AI chatbot, ahead of urgent care visits. Supervising physicians monitored the conversations in real time. According to Google, none needed to be interrupted under predefined safety criteria. The post does not list those criteria, so outsiders cannot judge how strict they were from this source alone.

Google also reports what clinicians and the system produced. Clinicians said AMIE's summaries helped them prepare for visits in 75% of cases and influenced their approach to care in more than half. The company says AMIE's differential diagnoses matched the doctors' final diagnoses 90% of the time. These are three different measures: preparation, influence on care, and diagnostic agreement. The blog post does not say how "matched" was defined, how many clinicians rated the summaries, or how many cases fall behind each percentage.

What is new is the setting. This was a real clinical environment, with real patients and a physician watching live, rather than a simulated exercise. What remains unconfirmed is almost everything beyond Google's own summary. This article draws on one source, a Google blog post written by a member of the research team, and we have not been able to read the paper itself or retrieve an independent account of the study. Treat the figures as the company's description of its own research.

Google itself cautions that larger clinical trials are needed to assess patient-facing AI at scale. It says the findings suggest potential to improve the patient-physician relationship and ease strain on healthcare workers. That is a statement of potential, not a demonstrated outcome.

The result matters because it bounds a question health systems are asking: whether AI can gather a patient's history before a visit without creating new risk. A 98-patient study at a single clinic, with clinicians supervising throughout, cannot show that an unsupervised system would behave the same way, or that results would hold at other sites or among other patients.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.