FDA Wants a New Kind of Proof for GenAI Medical Devices
Its discussion paper centers on competency testing, a two-axis risk framework and risk-scaled monitoring before any new rule exists.
Key Takeaways
- Competency, not a checkbox. The FDA is exploring a premarket assessment that pairs non-clinical benchmarking with clinical confirmation.
- Evidence across the lifecycle. The paper links risk assessment, premarket evaluation and postmarket monitoring instead of treating market entry as the end of the evidence story.
- The feedback window is open. Comments under docket FDA-2026-N-7874 are due October 19, 2026.
A GenAI-enabled medical device can produce new content, creating an evidence problem that conventional software review was not designed around. The FDA’s new discussion paper asks the practical question behind that problem: what should show that a generative system performs as intended before it reaches patients, and what evidence should follow it after deployment?
The agency has not proposed a binding rule. But its questions are concrete enough to matter now, especially for teams building devices around foundation models or agentic AI capabilities.
The most interesting idea is borrowed from medical training
The FDA calls it a possible premarket “competency assessment.” The concept is inspired, at a high level, by how physicians are trained and evaluated. For a GenAI-enabled device, the paper describes combining non-clinical device benchmarking with clinical confirmation to evaluate whether the system performs as intended before reaching patients.
That framing is notable because the FDA starts from the premise that GenAI-enabled devices may introduce risks that differ from traditional software and other AI-enabled medical devices. The evidence question is therefore broader than a benchmark score: how should non-clinical testing and clinical confirmation work together to demonstrate performance for a specific medical use?
Risk assessment is only the first layer
The discussion paper also outlines a possible two-axis framework for risk assessment and asks how postmarket monitoring should be scaled to risk. Foundation models and agentic AI systems appear separately in the paper as areas that deserve focused regulatory consideration.
Taken together, the questions point toward a lifecycle view of evidence: characterize the risk, evaluate performance before market, then keep measuring what happens after deployment. The FDA is asking for feedback on how rigorous each layer should be and what evidence belongs in it.
Why this matters before any rule exists
The paper is explicitly for discussion only. It is not draft or final guidance, does not change current policy and does not announce new submission requirements. That boundary matters.
What also matters is the direction of inquiry. The FDA is publicly testing ideas about competency, risk and postmarket evidence before it decides whether any of them should become formal regulatory expectations. Developers, clinicians and researchers therefore have a chance to challenge the assumptions while the framework is still being shaped.
The FDA is thinking beyond the U.S. market
This remains a U.S. regulatory consultation, and it changes no rule outside the United States. But FDA device-center director Michelle Tarver said the process could help inform an approach that serves as a potential model for regulators around the world.
For multinational medical-device teams, that makes the docket worth watching even before the FDA decides what comes next. The ideas the agency eventually chooses to formalize could influence how other regulators frame the same evidence problems.
What to watch by October 19
Manufacturers, clinicians, consumers, researchers and other interested parties can submit comments under docket FDA-2026-N-7874 through October 19, 2026. The immediate question is not whether the FDA has written the GenAI rulebook. It has not. The question is which ideas about risk, competency and postmarket proof survive public feedback, and whether any later appear in formal FDA guidance or regulatory expectations.
