FDA Signals GenAI Medical Devices May Face Physician-Like Competency Review, As Draft Policy Moves Ahead

The FDA’s emerging view of generative AI-enabled medical devices is starting to diverge from its long-running approach to software as a medical device. Speaking Tuesday at The MedTech Conference hosted by AdvaMed in Boston, Rick Abramson said these products may need to be assessed in a way that looks more like how physicians are evaluated for competency than how a conventional software tool is tested.

That distinction matters because, in Abramson’s description, generative AI changes both the input and output problem. Traditional software and earlier AI devices could often be framed in a bounded evidentiary model: put in a defined input, get a defined output. Abramson contrasted that with GenAI systems, where limitless text can go in and limitless text can come out, creating what he called an "unbounded evaluation space."

Why FDA Thinks The Old Software Model Breaks Down

Abramson said the agency has spent 20 years evaluating software as a medical device and is familiar with products that behave predictably enough to support straightforward evidence generation. He used radiology as an example, describing a classic AI device that takes a chest X-ray and returns a binary signal such as whether a pneumothorax is present.

For generative AI-enabled tools, he said, that model no longer holds. The problem is not simply that the technology is more advanced software. It is that exhaustive testing across every possible scenario is not practical or possible. Abramson drew a comparison to his own training and practice as a physician: over time he was evaluated on knowledge, clinical judgment, communication ability and adherence to safety boundaries, but he was never tested against every possible chest X-ray he could encounter over a career.

The regulatory signal is that FDA may be moving toward an assessment framework centered on demonstrated competence within constraints, rather than a product review model that assumes the full behavior space can be enumerated before clearance or approval. For developers, that would shift the compliance burden toward how systems are bounded, monitored and shown to operate safely in clinically meaningful conditions.

Draft Policy Is Coming After The August Discussion Paper

Abramson tied those remarks to the FDA’s August discussion paper on regulation of GenAI-enabled medical devices, which sought feedback from industry. He said the paper was not meant as a thought exercise and stated directly that the agency intends to go to policy and intends to draft policy.

The comment period ended on Monday, and Abramson said the agency has received more than 300 comments. He did not give a precise timeline for draft guidance, but he said it would arrive not on the course of weeks and certainly not on the course of years; instead, he said stakeholders should expect draft policy guidance on the course of months, followed by another opportunity for comment.

He also said this administration has made policy development in this area a priority, adding that he was referring to attention at the level of the HHS Secretary and the White House. That suggests the eventual framework could carry weight beyond a narrow technical guidance exercise and affect product planning across digital health and medtech.

The Industry Stakes

One reason this matters is that FDA is acknowledging a mismatch between legacy device review methods and GenAI behavior. Once a regulator says the existing evidentiary toolkit is insufficient, the commercial question becomes which companies can adapt fastest to a new proof standard.

The American Medical Association is already trying to shape that standard. Abramson referenced feedback coming into the discussion paper process, and the source notes that AMA CEO John Whyte sent a letter last week with recommendations on regulating generative AI and agentic AI tools. Whyte argued that neither the probabilistic nature of GenAI outputs nor the impossibility of testing every conceivable scenario should diffuse accountability across clinicians, patients, deploying institutions, foundation-model developers or other participants in the ecosystem.

That point lines up with the likely strategic tension ahead. If FDA adopts a more clinician-like competency model, it may create room for regulation that better fits adaptive and language-based systems. But it does not reduce the need to assign responsibility for failures. For developers, the opportunity is a pathway that recognizes how these systems actually function; the risk is that a more nuanced framework could still demand clearer accountability than many current deployment models provide.