Google Tested Its Own Medical AI On Real Patients
Google published a post on October 8, 2026 about a clinical study of AMIE, its research diagnostic AI chatbot, published the same day in The Lancet. Everything below comes from Google's own announcement. Nothing here is a product you can use: Google describes AMIE as a research system and says larger clinical trials are still needed.
Transcript
Google published a study on its medical AI. Nothing here is a product you can use yet.
The Lancet published it in the main edition, where Google had never appeared.
Ninety eight patients consulted AMIE before urgent care visits at a hospital primary care clinic. Supervising physicians watched in real time.
No conversation had to be stopped on the safety criteria. AMIE's diagnosis list usually matched what the doctors concluded.
Google helped write a study about its own system, and the group was small. Google says larger clinical trials are needed.
We read Google's blog summary, not the paper, and found no outside confirmation.
Inboxsmith answers your business calls with AI. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- HONESTY SPINE. What is new on October 8 2026 is that a study was published in The Lancet. No product launched, nothing is available, and AMIE is described on the page as "our research diagnostic AI chatbot", a research system that is nowhere offered to anyone. The second point, the editorial heart of this Short, is that the study was "led by researchers at Google and Beth Israel Deaconess Medical Center (BIDMC)" while the AI being evaluated is Google's own, so this is Google publishing a favourable result about Google's own system; that is provenance, not an accusation, and it is literally what the page says. Third, 98 patients is a small group and Google itself writes "While larger clinical trials are needed to assess patient-facing AI at scale". Fourth, the 90% figure is agreement with the doctors' final diagnoses, not proven correctness against any outside ground truth, and the page never states that AMIE outperformed any physician. Fifth, we read Google's blog summary and not the Lancet paper, so every figure is Google's own summary of its own work. EVIDENCE, CLAUSE BY CLAUSE. HOOK, "Google published a study on its medical AI." Backed by "A study led by researchers at Google and Beth Israel Deaconess Medical Center (BIDMC) and published in The Lancet today, offers an encouraging first look." and by "98 patients consulted AMIE, our research diagnostic AI chatbot". HOOK, "Nothing here is a product you can use yet." Backed by the verified absence noted in the source file, "no availability, pricing, rollout or launch date; AMIE is described as a research system", and by the page describing AMIE only as "our research diagnostic AI chatbot". This is the literal equivalent of the banned boilerplate sentence, which was deliberately not used. SCENE 1, "The Lancet published it in the main edition, where Google had never appeared." Backed word for word by "As Google's first-ever publication in the main journal of The Lancet, this research marks a significant milestone." and by "published in The Lancet today". The source's phrase "marks a significant milestone" is an idiom and was barred from narration; it appears only inside quotation marks on the scene 1 card, marked as the source's own wording. SCENE 2, "Ninety eight patients consulted AMIE before urgent care visits at a hospital primary care clinic." Backed by "In a real-world clinical study at BIDMC's ambulatory primary care clinic, 98 patients consulted AMIE, our research diagnostic AI chatbot, ahead of urgent care visits." The hospital is named only on the silent card because the TTS garbles that name; no city was invented, because the page never names one. SCENE 2, "Supervising physicians watched in real time." Backed by "Supervising physicians monitored the interactions in real time". SCENE 3, "No conversation had to be stopped on the safety criteria." Backed by "not a single conversation needed to be interrupted based on predefined safety criteria". SCENE 3, "AMIE's diagnosis list usually matched what the doctors concluded." Backed by "AMIE's differential diagnoses also matched the doctors' final diagnoses 90% of the time." The wording is deliberately qualitative and deliberately framed as agreement with the doctors rather than as accuracy, and the percentage is spoken nowhere. SCENE 4, "Google helped write a study about its own system." Backed by "A study led by researchers at Google and Beth Israel Deaconess Medical Center (BIDMC)" plus "AMIE, our research diagnostic AI chatbot" (the possessive "our" is Google's). SCENE 4, "and the group was small." Backed by the published size, 98 patients, and by Google's own call for larger trials. SCENE 4, "Google says larger clinical trials are needed." Backed verbatim by "While larger clinical trials are needed to assess patient-facing AI at scale", using the source's own hedge. SCENE 5, "We read Google's blog summary, not the paper, and found no outside confirmation." Backed by the provenance notes in the source file: "We read Google's blog post. We did NOT read the underlying Lancet paper, so every figure here is as Google summarised it on its own blog." and "0 of 5 corroborating YouTube transcripts retrieved". ENDCARD. Inboxsmith pitch plus the exact required closing sentence; no claim about the story. EXCLUDED AND WHY. The 75% of cases in which summaries "helped them prepare for visits", the "more than half" who said summaries "influenced their approach to care", the 90% diagnosis agreement, the zero interrupted conversations as a digit, and all dates: all excluded from speech under the one-spoken-figure rule and carried only as digits on the silent cards. The hospital name and the acronym BIDMC: excluded from speech as TTS hazards, carried on the cards. The 10 timeline probes: discarded as unrelated noise, cited nowhere. The "Related stories" cards: separate posts, evidence for nothing, excluded. Our own earlier Short about a different AMIE study with patient actors: not referenced, and the narration instead makes clear that what is new is this published study. The Lancet paper itself: not read, so no claim rests on it. Any suggestion of a launch, availability, rollout, pricing, or of AMIE beating a physician: excluded because the page publishes none of it. GATE NOTES. DASHES: no em-dash and no en-dash appears in any field of this file; separation is done with commas, full stops and parentheses, and the only hyphens present sit inside verbatim quoted source wording (for example "patient-physician", "real-world", "first-ever", "patient-facing", "p-values", "On-page", "yt-dlp"), never in narration. All quotes and apostrophes are straight ASCII. BANNED WORDS AVOIDED: "agentic" appears nowhere; the sentence "There is no product launch here." was not used and was replaced by the literal equivalent "Nothing here is a product you can use yet."; "expires" and any countdown or deadline phrasing appear nowhere; no brand name is glued to digits in narration; "launched", "rolling out", "available" and "announced a product" appear nowhere in narration. FIGURES: exactly one figure is spoken in the entire script, "ninety eight patients" in scene 2, written as words, and it is spoken exactly once; no other digit, percentage, number word or date is spoken in the hook, in any scene narration, or in the endcard. LISTS: no narration sentence contains a multi-item list or an "A, B, and C" enumeration; every enumeration lives on the silent cards. DATES: no date is spoken anywhere; dates appear only as digits on the silent cards and in this ledger. IDIOMS BARRED: "step into a doctor's office", "push the boundaries of what's possible", "an encouraging first look", "ease the strain", "marks a significant milestone" and "at scale" are all barred from narration; in their place the narration says the literal things, that a study was published, that patients consulted AMIE before urgent care visits, that no conversation had to be stopped, that the diagnosis list usually matched what the doctors concluded, and that Google says larger clinical trials are needed. TIMELINE: no timeline array is included, on purpose, because this story is same-day and a timeline block would age it into the catalog tier and break the freshness wording. COUNTS: narrated word count 118 (hook 17, scene 1 13, scene 2 21, scene 3 19, scene 4 21, scene 5 13, endcard 14); sentence count 12 (hook 2, scene 1 1, scene 2 2, scene 3 2, scene 4 2, scene 5 1, endcard 2); every scene narration is at or under the 30 word per scene cap; the hook is exactly two sentences and 17 words; planning durationS values sum to 40, inside the 30 to 50 window.(First-party corporate blog post published by Google on its own host, blog.google, in the Health section. The page title reads "Study in The Lancet suggests AI could improve patient-physician relationships." and the on-page headline reads "Study: AI could improve patient-physician relationships". The byline is Mike Schaekermann, with the stated role Research Lead, and the on-page date is Oct 08, 2026, the same day as this production run. The live URL was fetched successfully on 2026-10-08 and the body matches our watcher snapshot captured at 2026-10-08T22:32:18.874Z verbatim, so there is no drift between what the watcher saw and what is live now. The snapshot is stored at scripts/data/article-cache/9412e609363df4d9.txt. This is the publisher's own domain writing about the publisher's own research result, not an aggregator, not a syndicated reprint, and not a link-out stub to a third-party host, so provenance is clean first-party, with the important caveat that the subject of the study is Google's own AI system and Google is a co-author of the study. Corroboration status is zero: 0 of 5 corroborating YouTube transcripts were retrieved because this host's yt-dlp build cannot fetch captions, which is a known local tooling fault rather than an absence of coverage, and the 10 dated timeline probes returned unrelated noise (statin videos, vitamin D videos, foreign-language drama shorts about doctors) that corroborate nothing, so they were discarded entirely and appear nowhere in the narration or the on-screen cards, not even by implication. The page's "Related stories" cards (Google Earth AI public health, MedGemma, MrBeast and Gemini, Building the future of global health) are separate posts and were used as evidence for nothing in this script. The page links out to AMIE, to "AI in science and health", to Beth Israel Deaconess Medical Center and to the complete research paper in The Lancet; we did not read the paper itself. No prompt-injection or instruction-like content was found anywhere in the source text, which reads as ordinary corporate research communications prose.)