Diagnosis (general) · AI doctors and chatbots
AMIE: Google's research chatbot tested with real patients before GP visits
In a Lancet study published 8 October 2026, 100 adults at one Boston primary-care clinic text-chatted with Google's research AI, AMIE, before an urgent appointment, and 98 then had the visit. A doctor watched every chat live, and none had to be stopped under the safety rules. It was a small, Google-funded feasibility study with no comparison group.
How far away is this?
This is the first published study we've found of AMIE with real patients, but it was a small feasibility study, not a trial of benefit. There was one clinic, no comparison group, about 100 English-speaking patients who had to use a laptop or desktop, and a doctor supervising every chat. Patients had already been triaged as not needing emergency care, and pregnancy and mental-health visits were excluded, so emergency and psychiatric triage weren't tested. It doesn't show that AMIE improves patients' health or is safe without supervision.
Where is it approved?
| Region | Regulator | Status | Notes |
|---|---|---|---|
| United States | FDA | Not approved | Research system, not offered to patients. Registry: not an FDA-regulated device study. No openFDA 510(k) record found (checked 16 Oct 2026).[3] |
| European Union | CE mark / EMA | Not approved | Research system, not offered to patients (Google calls it a research chatbot). We found no CE mark; EUDAMED not searched.[5] |
| United Kingdom | MHRA | Not approved | Research system, not offered to patients (Google calls it a research chatbot). We found no MHRA registration; PARD not searched.[5] |
| Australia | TGA | Not approved | Research system, not offered to patients (Google calls it a research chatbot). We found no ARTG entry; the ARTG search was unavailable.[5] |
A research system, not offered to patients anywhere. 'Not approved' is based on that and on finding no regulator records, not on a regulator statement. "Unknown" means we haven't verified it, not that it's unavailable. Approval in one country isn't approval in another.
What is it?
AMIE is a research chatbot built by Google. It asks about your symptoms, like a doctor taking your history. In this study it ran a text chat a few days before the appointment, then suggested possible diagnoses for the patient to discuss with their clinician. The clinician got a summary and the transcript. AMIE's suggested management plans were kept for research only.
What's the evidence?
Prospective, single-centre, single-arm feasibility study at Beth Israel Deaconess Medical Center (Boston). English-speaking adults chatted with AMIE up to 5 days before a single-complaint urgent primary-care appointment, while physician safety supervisors watched.[2]
Evidence stage: In clinical trials · #c1
114 patients enrolled between April and November 2025. 100 completed the AMIE chat, and 98 completed both the chat and the clinic visit.[4]
Evidence stage: In clinical trials · #c2
Primary outcomes: safety stops, conversation quality rated by clinical evaluators, and patient and clinician surveys. There were zero safety stops under the pre-agreed criteria. Supervisors noted one hallucination and added clinical information in five chats. Evaluators rated the conversations favourably in 87–100% of cases across 17 criteria; patients did so in 48–96% across 16 criteria.[2]
Evidence stage: In clinical trials · #c3
Clinicians filled in post-visit surveys for 60 of 98 visits; none for the other 38. In 16 of the 60 the clinician hadn't read the transcript. Of the 44 who had, 33 found it helpful and 25 said it might have changed what they did.[4]
Evidence stage: In clinical trials · #c4
Secondary outcome (not a primary one): the authors' March 2026 preprint reports that the final diagnosis, from a chart review 8 weeks later, appeared in AMIE's list of up to 7 possible diagnoses in 88 of 98 cases (90%), counting near-misses. It was in the top 3 in 75%, and AMIE's first guess matched in 56%. In 52 of 98 cases the 'final diagnosis' was the clinician's judgement without a confirming test. Blinded reviewers found no significant difference from the clinicians' diagnosis lists, but rated the clinicians' management plans as more practical and cost-effective. The preprint isn't peer-reviewed, the Lancet abstract doesn't report these figures, and we haven't checked them against the final paper.[4]
Evidence stage: In clinical trials · #c5
Funded by Alphabet. Many authors are Google employees who hold Alphabet stock and Google-related patent interests, and the senior author was a visiting researcher at Google. The registry lists Beth Israel Deaconess Medical Center as lead sponsor and Google as collaborator.[2]
Evidence stage: In clinical trials · #c6
The registry lists it as an observational cohort with no data monitoring committee, not an FDA-regulated device study, and a recruitment target of 100–200 (actual enrolment recorded as 100). Its 2025 summary planned to withhold AMIE's diagnoses and plans from patients and clinicians; the authors' preprint reports that the diagnoses were shown (see c8). A supervising doctor debriefed each patient after the chat to correct any errors. Visits for pregnancy or mental-health concerns were excluded.[3]
Evidence stage: In clinical trials · #c7
Patients saw AMIE's list of possible diagnoses at the end of the chat, framed as things to discuss with their clinician, and clinicians saw it in the transcript. AMIE's management plans weren't shown to anyone. After 50 chats, the underlying model was switched from Gemini 2.5 Pro to the faster Gemini 2.5 Flash. Patients had to use a laptop or desktop and be on the clinic's online portal. The authors note patients knew they were being observed, which may have inflated ratings (the Hawthorne effect).[4]
Evidence stage: In clinical trials · #c8
What does it mean for me?
You can't use AMIE; it's a research system. The study suggests a closely supervised AI history-taker can be acceptable to patients and clinicians. It doesn't show that unsupervised chatbots are safe. If a chatbot tells you something is minor but you feel very unwell, seek care anyway.
What could I ask my doctor?
- Is any AI used to gather my history before an appointment?
- Who reads what I tell the AI, and is a person checking it?
These are general starting points, not a recommendation to choose any particular test or treatment.
Change history
- : First published.
- : Review label corrected: this page was not clinically reviewed. Earlier labels saying 'independent fact-check' and 'clinical review' were wrong; checks so far were AI-assisted source checks only.
General information only, not medical advice. Talk to your doctor about your situation. AI-assisted source check. Not clinically reviewed.
References
- Brodeur et al., The Lancet (2026), DOI. doi.org (2026). Peer-reviewed study. https://doi.org/10.1016/S0140-6736(26)01535-7
- PubMed abstract and declaration of interests (PMID 42849491). pubmed.ncbi.nlm.nih.gov. Abstract. https://pubmed.ncbi.nlm.nih.gov/42849491/
- ClinicalTrials.gov NCT06911398. clinicaltrials.gov. Trial registry. https://clinicaltrials.gov/study/NCT06911398
- Authors' preprint, arXiv:2603.08448 (March 2026). arxiv.org (2026). Preprint (not peer-reviewed). https://arxiv.org/abs/2603.08448
- Google blog announcement. blog.google. Company document (claim). https://blog.google/innovation-and-ai/technology/health/amie-clinical-study-lancet/
- Google Research blog. research.google. Company document (claim). https://research.google/blog/exploring-the-feasibility-of-conversational-diagnostic-ai-in-a-real-world-clinical-study/
Publisher shown as the website hosting each source. Dates are given where the source label includes them.