AMIE in The Lancet: promising, supervised, not yet proven
The most closely watched result of the month came on October 8, when The Lancet published a study of Google's conversational diagnostic system, AMIE, run with Beth Israel Deaconess Medical Center in Boston. It was Google's first paper in the main Lancet journal.
Patients who had booked an urgent primary care visit chatted with AMIE from home, up to five days beforehand, while doctors watched every conversation in real time. Of 98 completed conversations, none had to be stopped under the study's safety criteria, and supervisors flagged one hallucination. AMIE's list of possible diagnoses included the final diagnosis in 90 percent of cases, ranked it in the top three 75 percent of the time, and put it first in about 56 percent. Blinded reviewers rated its differentials as comparable to the primary care doctors', though the physicians did better on whether a management plan was practical and cost-effective. Doctors who read the AMIE summary said it helped them prepare in about three quarters of those cases.
It is an encouraging result, and the authors are careful about it. It was a single-site feasibility study with no control group, a text-only interface, English-speaking patients and a sample that skewed younger than the clinic's usual patients. Alphabet funded it, and one author was a visiting researcher at Google during the study. One physician noted a case where listing lymphoma among the possibilities might have caused a patient needless worry. The researchers were explicit that this shows the system can operate safely under supervision, not that it improves outcomes.
That distinction is the heart of the matter. "Safe under supervision" and "safe on its own" are different claims, and a large part of the policy argument in healthcare AI is really an argument about which of the two a given product has earned. Take away the supervising physician, add a second complaint, or move to a different kind of clinic, and the result no longer applies.
Imaging is the mature story
If you want to see healthcare AI at scale, look at radiology. As of September 2025, the FDA had authorized roughly 1,350 AI-enabled medical devices, and about three quarters were radiology tools. These are narrow systems doing a defined job, such as flagging a suspected bleed or a lung nodule, and a radiologist still reads the scan.
On October 1, a dozen health systems joined a diagnostic AI collaborative organized around the imaging company Aidoc. It is a small but telling sign that hospitals increasingly want to evaluate these tools together, sharing results, instead of one hospital at a time.
Drug discovery is finally facing its exam
For years the pitch was that AI would find drugs faster and cheaper. In 2026 the bill is coming due, because candidates are reaching the stages where a drug either works or it does not.
The headline case is Insilico Medicine's rentosertib for idiopathic pulmonary fibrosis. The company announced the start of a Phase III trial in July, widely described as the first Phase III study of a generative-AI-driven drug, and has since reported dosing its first patient. Isomorphic Labs, the Alphabet spinout built on AlphaFold, has been preparing for human trials of its own, with oncology first.
But as of October, no drug discovered or designed by AI has received full FDA approval. A published analysis cited in industry reviews found AI-discovered drugs failing in Phase 2 at about the same rate as conventionally discovered ones. AI seems good at getting molecules into the clinic faster. Whether it gets more of them across the finish line is the open question, and the Phase III results over the next couple of years are where it gets answered.
There is also a quiet fight over labels. Takeda's zasocitinib beat an approved rival in a head-to-head psoriasis trial in June, and the FDA accepted its application in September. Whether it counts as an "AI drug" is disputed, since the work started years ago against an already-validated target. When a label is worth billions in investor attention, people stretch it. It is worth remembering whenever a company says its product is "AI-powered."
What the evidence says so far
Step back and the picture is consistent across both areas. AI is genuinely useful when a clinician or scientist is in the loop and the task is well defined. The leap to doing the job unsupervised hasn't been shown yet, in diagnosis or in drug development. That gap is where the rules come in, and the rules are the subject of the final article.
