FDA opens rulemaking discussion for generative AI medical devices

The US Food and Drug Administration has opened a public discussion about how generative artificial intelligence in medical devices should be regulated. On August 18, the agency published a discussion paper and invited manufacturers, clinicians, researchers, consumers and other interested parties to respond by October 19, 2026. The initiative does not create binding rules yet. It begins a formal evidence-gathering process that could shape future requirements for products that generate text, images, recommendations or other clinical outputs. The FDA says these systems can differ from traditional software and earlier AI-enabled devices because their outputs may be open-ended, their behavior can depend on foundation models and some products may act through agentic workflows. Its paper proposes a possible two-axis framework for assessing risk. It also explores a premarket evaluation model based on competency assessment, loosely inspired by how physicians are evaluated. Under that concept, developers could need to demonstrate performance through non-clinical benchmarking and clinical confirmation before a device reaches patients. For products already on the market, the agency is considering risk-proportionate monitoring to detect performance changes and emerging problems. The paper also asks how responsibilities should be divided when a device maker builds on a foundation model supplied by another company. This matters because medical AI developers currently face uncertainty about what evidence will satisfy regulators when systems produce variable responses or change over time. Clearer expectations could affect product design, documentation, testing costs and release schedules. Clinicians and patients also have a direct interest: useful generative systems may support care, but unreliable or poorly monitored outputs can create safety risks. The FDA frames the consultation as an attempt to protect patients while allowing responsible innovation. That balance is still a proposal, not a settled policy, and the final approach may change after public feedback. For AI makers, the immediate practical step is to study the questions in the paper and document how their systems are benchmarked, clinically confirmed and monitored throughout their lifecycle.