ARTICLE SUMMARY:
In this week’s Pathways Picks: FDA proposes a “competency” framework for regulating generative AI; Trump is expected to nominate White House aide and MAHA proponent to FDA; TEMPO selections continue; RAPID proposal released; device reclassifications; Singapore cybersecurity guide; Brazil’s innovation pathways; and more global medtech policy updates.
FDA Floats GenAI Plan
Generative AI technologies seeking medical device authorization would have to demonstrate “competency” against a series of non-clinical benchmarking tests followed by some type of clinical confirmation under a framework proposed by FDA in an August 18 discussion paper. The approach, the agency suggests, is intended to draw on a series of recent peer-reviewed recommendations that increasingly autonomous generative AI tools, including large language models (LLMs), be regulated more like human practitioners than medical devices, but “adapted for the technical, practical, and legal considerations applicable to the regulation of medical devices.”
FDA has yet to authorize generative AI devices (with one possible exception of a highly restrained LLM), but has taken a series of steps in recent years, including 2024 and 2025 advisory committee meetings, to build a foundation for doing so, despite the fact that these tools are inherently more freewheeling and less readily transparent in how they work. “CDRH recognizes that evaluation approaches developed for software with bounded inputs and fixed outputs may not be appropriate for GenAI-enabled devices,” the device center states.
Under the proposed framework, developers seeking authorization would need to put their technology through a series of agreed-upon benchmarking tests to provide a point-in-time assessment of how the deployed AI system will function in the real world, followed by some type of either retrospective or prospective clinical assessment. They would then need to repeat the benchmarking periodically in the postmarket phase, combined with periodic clinician review and performance-degradation monitoring.
FDA has used discussion papers like this one as a first step to developing official guidance and regulatory policies. It took a similar approach, for instance, to implement policies for non-generative AI tools and the concept of predetermined change control plans (PCCPs). Public comments on the draft are due October 19.
