A lead review published in Frontiers in Science on 1 September, by authors from Imperial College London, Stanford, Yale and Harvard Law School, surveys where medical AI has actually landed. Its central observation is that the FDA's AI-Enabled Medical Devices List holds more than 1,500 entries while almost none of that technology has reached daily clinical practice.
What the common framing gets wrong
Three corrections matter before the number is quoted. First, those 1,500-plus entries are overwhelmingly 510(k) clearances — findings of substantial equivalence to an existing marketed device — not FDA approvals. The review's own phrase “regulatory-approved” is the standard slippage, and the distinction is the whole difference between a device shown to be safe and effective and a device shown to resemble one already sold.
The trial result is non-inferiority
Second, the headline randomised evidence is a trial of over 80,000 women in mammography screening — not chest radiography, as it is sometimes rendered — and the finding is that AI-supported screening was non-inferior to standard double reading. AI matched two radiologists; it did not beat them. Non-inferiority is a real and economically significant result, and it is not the result usually reported.
A review is not a study
Third, this is a review article. No new data were collected. Coverage reporting that a “new study finds AI can predict 3,000 diseases” is quoting a citation inside a literature survey: the biobank-trained models predicting over 3,000 diseases undiagnosed at recruitment, and the Delphi-2M transformer modelling trajectories across more than 1,000 diseases, are other people's papers being summarised. One further figure is worth dating: the 73% of healthcare cybersecurity professionals reporting outdated operating systems comes from a 2021 survey, five years stale.
Why a deflationary review is the useful one
The authorship is unusually senior and unusually cross-disciplinary, pairing clinical and machine-learning researchers with a Harvard Law health-policy scholar. Its claim — that the clinically mature, regulator-cleared AI in hospitals today is almost entirely conventional deep-learning diagnostics, and that the generative wave has not entered routine care — is a direct check on the framing that AI is already transforming medicine.