FDA Weighs Treating Generative AI Medical Tools Like Physicians, Not Devices

When your doctor orders a chest X-ray to check for a collapsed lung, you trust that the person reading it has been trained, tested, and held accountable over years of practice. Nobody expects a physician to have personally reviewed every possible X-ray that could ever exist before earning that trust. Instead, we evaluate their judgment, their knowledge base, and their ability to recognize when something falls outside their expertise.

That same logic may soon apply to artificial intelligence tools used in medicine, according to a senior FDA official. Speaking Tuesday at The MedTech Conference in Boston, an event hosted by the industry group AdvaMed, Rick Abramson, a Harvard-trained radiologist who directs the FDA's Digital Health Center of Excellence, said generative AI tools might need to be assessed less like a piece of software and more like a clinician whose competency is monitored over time.

This is a meaningful shift. For roughly two decades, the FDA has evaluated software as a medical device using a fairly predictable framework. You feed in a specific input, the software produces a specific output, and regulators can test that relationship against known cases. Abramson used his own field as an example: an AI tool analyzes a chest X-ray and returns a binary signal, flagging whether a pneumothorax, a collapsed lung, is present. That kind of system is bounded. You can test it against thousands of known X-rays and get a reasonably clear picture of how reliable it is.

Generative AI breaks that model. These tools can accept open-ended text as input and produce open-ended text as output, creating what Abramson called an "unbounded evaluation space." There is no finite library of possible inputs to test against, and no simple yes-or-no answer to validate. The traditional evidentiary playbook, built for software that behaves predictably within narrow boundaries, simply was not designed for tools that can generate novel responses to novel situations.

"We don't have the methods for that type of evidentiary proof," Abramson told the audience.

A New Evaluation Model Takes Shape

So what does the agency propose instead? Abramson pointed to an idea the FDA floated in an August discussion paper: treating these generative AI systems more like licensed professionals than like static devices. He drew directly from his own career as a physician to explain the concept. Over the course of his training, Abramson said, he was evaluated on his medical knowledge, his clinical judgment, his ability to communicate with patients, and his adherence to safety boundaries. At no point was he tested against every conceivable X-ray he might ever encounter in practice.

"This is not practical or possible," he said.

That comparison matters because it reframes the regulatory question. Instead of asking "does this software produce the correct output for every possible input," the FDA may start asking something closer to "does this system demonstrate sound judgment, communicate appropriately, and recognize the limits of its own competence." It is a more holistic standard, similar to board certification or ongoing competency review for doctors, rather than a one-time pass-fail test on a fixed dataset.

The FDA's August discussion paper was designed to gather industry feedback on exactly this question: how should generative AI-enabled medical devices be regulated going forward. The comment period closed this past Monday, and the agency has already received more than 300 responses. That is a substantial volume of input, reflecting how much is riding on getting this framework right for device makers, hospitals, clinicians, and ultimately patients.

When AdvaMed's vice president of technology and regulatory affairs, Geeta Pamidimukkala, who moderated the panel, asked what comes next, Abramson was unambiguous that this work is headed toward formal policy. "We intend to go to policy. We intend to draft policy. This is not just an academic paper," he said. He added that the push has backing well above his own office. "This administration has made this policy development a priority, and I'm talking about at the level of the [HHS] Secretary and the White House."

On timing, Abramson was deliberately noncommittal, though he offered a rough window. The agency wants to actually read and synthesize the 300-plus comments rather than rush the process. "We're not just going to run them through an AI engine," he said, drawing laughter from the audience. His estimate: not weeks, not years, but likely months before draft guidance emerges, followed by another round of public comment.

Not everyone is convinced the physician analogy resolves every concern. The American Medical Association weighed in with its own letter to Abramson last week, signed by CEO John Whyte, offering recommendations for how the FDA should build out its framework for generative and agentic AI tools. One theme Whyte raised is practicality, the same issue Abramson says is reshaping the agency's thinking. But Whyte also pushed back on a risk embedded in the new approach: the idea that because generative AI is probabilistic and impossible to test exhaustively, accountability might get diluted. "Neither the probabilistic nature of GenAI outputs nor the impossibility of testing every conceivable scenario should diffuse that accountability across clinicians, patients, deploying institutions, foundation-model developers, or other participants in the ecosystem," he wrote.

That tension, between a flexible evaluation model and clear lines of responsibility, is likely to shape much of the public debate as the FDA moves from discussion paper to formal draft guidance.

Why It Matters

This shift in thinking carries real stakes for patients and the clinicians who rely on these tools daily. A regulatory framework that treats generative AI like a licensed professional, subject to ongoing review rather than a single upfront test, could better reflect how these systems actually behave in practice. But it also raises hard questions about who bears responsibility when something goes wrong, a question the AMA is right to press. As the FDA moves from discussion paper to formal policy in the coming months, the balance it strikes between flexibility and accountability will shape how safely generative AI gets integrated into American healthcare for years to come.