Google’s AMIE Primary Care Feasibility Study Published in The Lancet

A prospective clinical study led by researchers at Google and Beth Israel Deaconess Medical Center evaluating Google’s AMIE conversational diagnostic AI in a real-world urgent primary care clinic was published in The Lancet on October 8, 2026, according to Google’s announcement. In the study, 98 patients consulted the AMIE chatbot ahead of urgent care visits while supervising physicians monitored the interactions in real time, and none of the conversations had to be interrupted under predefined safety criteria.

Google described the paper as its first-ever publication in the main journal of The Lancet. The company said larger clinical trials are needed to assess patient-facing AI at scale, and said the findings suggest potential for AI to enhance the patient-physician relationship and ease strain on healthcare workers.

Study Design and Safety Oversight

The system, formally named the Articulate Medical Intelligence Explorer (AMIE), is a medical AI chatbot that patients used from home after scheduling an urgent primary care appointment with a BIDMC physician. According to the Google Research team’s write-up, 100 adult patients completed a pre-visit interaction with AMIE, and 98 attended their scheduled appointment. Participants, who were booked for new, non-emergency, episodic complaints, interacted with the system through a secure text-chat interface up to five days before the visit. The system asked about symptoms, gathered medical histories, presented potential diagnoses for patients to discuss with their doctor, and produced a summary that the clinician could review beforehand. The prospective, single-arm, single-center study was pre-registered on ClinicalTrials.gov and conducted under institutional review board-approved protocols, with patients assured that their participation decision would not affect their care.

Participants skewed younger than the clinic’s overall urgent care population: of 1,452 total urgent care visits during the study period, more than half involved patients over age 60, while the sample’s female and white population trends were consistent with the clinic’s population, the Google Research account states.

BIDMC reported that the study ran from April through November 2025 and enrolled 114 patients, and that every conversation was monitored in real time by a board-certified internal medicine physician via live video call with screen-sharing. Supervisors were trained to intervene under four pre-specified criteria: immediate concern for harm to self or others, significant emotional distress related to the AI interaction, supervisor-identified potential for clinical harm, or an explicit patient request to end the session. Across all 98 completed encounters, no conversation required a safety stop; supervising physicians identified one hallucination and provided additional clinical clarification in five cases, according to BIDMC’s October 9, 2026 release. “This study helps establish the baseline characteristics of such real-world conversations,” co-first author Peter Brodeur, a clinical fellow in cardiovascular medicine at BIDMC, said in the release. BIDMC said it believes the study is the first prospective real-world study of a patient-facing conversational AI system in primary care.

Diagnostic Reasoning Results

AMIE’s differential diagnosis included the final diagnosis, established through chart review eight weeks after each encounter, within its top seven possibilities in 90% of cases, with 75% top-three accuracy, and it identified the final diagnosis as its single most likely possibility in 56% of cases, the research team reported. Accuracy remained high for the 46-patient subset whose final diagnosis was confirmed by a diagnostic test such as a laboratory, microbiological, pathological, or imaging result.

In a blinded and randomized comparison, panels of three clinical evaluators per case rated differential diagnoses and management plans from AMIE and from primary care providers, finding similar overall quality for differential diagnoses (p = 0.6) and for the appropriateness (p = 0.1) and safety (p = 1.0) of management plans, per the study’s preprint record. The physicians outperformed AMIE on the practicality (p = 0.003) and cost-effectiveness (p = 0.004) of management plans. Google Research attributed that gap to AMIE’s lack of access to the patient’s electronic health record, its inability to perform a physical exam, and its lack of multimodal input such as a patient’s overall physical appearance.

Patient and Clinician Experience

Patients completed the General Attitudes towards AI Scale before the chat, after the chat, and after the provider visit; attitudes shifted significantly more positive after interacting with AMIE (p < 0.001) and remained elevated after seeing the provider, the researchers reported. BIDMC said patients rated the system highly for listening, explaining information, and helping them feel at ease, while concerns remained around trust in the confidentiality of information shared with the system and trust in the chatbot’s honesty and trustworthiness. Adam Rodman, Director of AI Programs at BIDMC’s Carl J. Shapiro Institute for Research and Education, said future research will need to explore which interaction characteristics can build patient trust and how AI interactions can enhance the patient-physician relationship.

Primary care providers reviewed an AI-generated transcript or summary before seeing patients in 44 cases; physicians said doing so helped them prepare for the visit in 75% of cases and may have influenced their clinical approach in 57%, BIDMC reported. In one case, a provider described the interaction as somewhat harmful, citing concern that a patient may have experienced anxiety after AMIE included lymphoma among its possible diagnoses. In qualitative interviews reported by Google Research, physicians said AMIE shifted the visit dynamic from simple data gathering to data verification, allowing more collaborative conversations and shared decision-making.

Limitations, Funding, and Prior Disclosure

The research team describes the work as a single-center feasibility study without controlled comparisons that does not support quantitative efficacy claims against a baseline workflow, and lists further limitations including the text-only interface and unexplored effects of health literacy, tech literacy, and chatbot familiarity. Rodman and colleagues emphasized the study was designed to evaluate feasibility rather than to determine whether AI improves health outcomes.

BIDMC disclosed that the study was funded by Alphabet and that Rodman served as a visiting researcher at Google during part of the study. Brodeur and Jacob M. Koshy are the paper’s co-first authors, Rodman is listed last among the authors, and Marc L. Cohen is co-senior author.

The study results were first publicly disclosed in March 2026: the preprint was submitted to arXiv on March 9, 2026, and the Google Research team published its detailed account on March 11, 2026, updating that post on October 8, 2026 to note the Lancet publication. The research team said it intends to keep assessing the utility and impact of such systems in forthcoming larger studies with controlled comparisons.