AI medical diagnosis regulation: FDA weighs a competency path as the agency faces pressure

US regulators are moving toward a competency-based approval route for generative AI medical devices, according to a Cooley analysis of an FDA proposal. The change matters because it could let AI diagnostic tools clear review on demonstrated performance rather than on the hardware they replace. That shift would reshape how such products reach clinics.

That proposal sits alongside a more visible signal of the FDA's direction. On Thursday 24 September, an FDA panel voted in favour of Grail's Galleri test, recommending it for premarket approval, the South China Morning Post reported. It was the first time US regulators endorsed a multi-cancer early detection blood test, which screens healthy, asymptomatic people for abnormal molecules including circulating tumour DNA. The endorsement is a regulatory milestone. It is not a clinical one.

Hong Kong doctors remain cautious about the test's clinical value and say more data is needed before wider adoption, the SCMP reported.

That gap between a regulatory green light and clinical acceptance is the central tension in AI-assisted diagnosis right now. Agencies are being asked to rule on tools whose evidence base is still thin. Clinicians are being asked to trust outputs they cannot fully audit. The FDA's competency-based path, flagged in a 28 September Cooley post, would route generative AI-enabled devices through review criteria tied to how the model performs its task. The agency had earlier opened a comment period on a framework for generative AI-enabled medical devices, according to a 29 September JDSupra entry. Neither document is final, and the practical effect depends on how the agency writes performance thresholds that vary by clinical use.

Evidence arrives faster than rules

Oncology is where the pressure is highest. Medical Daily reported on 1 October that an early Mayo Clinic AI model flagged pancreatic cancer risk up to three years before diagnosis by reading health records. That result would be impossible to replicate with a single scan or lab test. The same day, the-scientist.com described a platform giving oncology researchers access to AI cancer tools. Both point in the same direction: risk prediction is moving from the imaging suite into the electronic record, where the regulatory questions are different.

Regulators are not the only ones improvising. Hospitals and clinicians are making adoption decisions with limited guidance.

A Cureus pilot study published on 1 October surveyed junior medical officers in Sydney on their knowledge of and attitudes toward AI. It is a reminder that the people expected to act on model outputs are often the least trained in them. MedCity News argued on 30 September that the real constraint on AI in healthcare may be trust rather than autonomy. That is a different problem than the one device reviewers are built to solve. Into that vacuum has stepped politics. At an industry-backed MAHA Summit, Health Secretary Robert F. Kennedy Jr. said AI is "better informed" than doctors, according to The New York Times on 29 September and Forbes the same week. The International Business Times UK reported that Kennedy said Sam Altman had called it "malpractice" for doctors to diagnose or prescribe without checking AI. Those are political claims, not regulatory ones, and they carry no weight in an FDA review. But they set the tone for the public argument about who, or what, gets to make a diagnosis.

Some of that argument is playing out at the level of liability. expresshealthcare.in published a 30 September piece on the shift from "human-in-the-loop" to "human-on-the-hook," the idea that a clinician who accepts a model's recommendation absorbs the blame when it is wrong. That framing matters for regulation because it decides what hospitals will actually deploy. A tool with a clear audit trail and a defensible performance claim is easier to sign off on than a black box, regardless of which one is more accurate.

The evidence base itself is contested.

A study circulated in May 2026 asked whether AI really beat emergency room doctors at diagnosis and what the results actually showed. The answer was more qualified than the headlines suggested. More recently, Bioengineer.org reported on 30 September that privacy-preserving AI can learn personalised treatment rules without exposing patient data. That technical advance also changes the regulatory calculus, since the training data may never leave the hospital.

Beyond Washington

The United States is not writing these rules alone. Open Access Government reported on 30 September on GUIDE-AI, a European initiative to bring guideline-directed medical treatment into the age of AI. Medical Buyer reported on 30 September that a new report recommends Britain modernise its medtech regulatory regime. ukauthority.com reported the same week that the MHRA is boosting health innovation. Earlier in September, a proposal for "L-plate" AI healthcare regulation in the UK signalled a graduated approach rather than a single gate.

In Asia, the picture is more fragmented. Saudi Arabia's SFDA approved Saudi-developed AI-enabled medical software for dental and ophthalmic diagnosis. India's CDSCO issued guidance clarifying the regulatory pathway for AI and software-based medical devices. Both were reported in September. The WHO's 2021 guidance on AI in health remains the closest thing to a common reference, and it is not binding.

Payment is the quieter lever. STAT reported on 1 October that Blue Shield of California, one of 17 major health plans committed to adopting payment models inspired by Medicare's ACCESS Model, expects to launch similar models in commercial plans in 2027. The insurer's chief medical officer, Ravi Kavasery, told STAT the company is still working out details including whether it will adopt CMS payment amounts. Whatever the FDA approves, tools that do not fit a reimbursement code tend not to reach patients at scale.

The most concrete near-term question is what the competency-based path actually requires. If the FDA asks for prospective trials, few generative diagnostic tools will clear it soon. If it accepts retrospective performance data, approvals could come quickly and the burden shifts to clinicians and payers to decide what evidence is good enough. The agency has not said which way it will go, and the comment period on its earlier framework closed without a published rule.

What is clear is the direction of travel. A year ago, multi-cancer blood tests were a research story. As of 24 September, one has an FDA panel recommendation. The regulatory scaffolding around AI diagnosis is being built in public, in pieces, and mostly after the tools are already in use.