Source: FDA — 2026-08-18
Summary
The FDA's Digital Health Center of Excellence issued its first discussion paper specifically addressing generative AI-enabled medical devices — the class of tools built on large language models and foundation models rather than traditional locked algorithms. The paper proposes a two-axis risk framework, a premarket evaluation approach built on competency assessment, and a plan for risk-proportionate postmarket monitoring, and explicitly flags foundation models and agentic AI systems as needing new regulatory thinking. Feedback is open under docket FDA-2026-N-7874 through October 19, 2026, and the agency stresses this is not yet proposed policy.
Key Takeaways
- This is the FDA's first dedicated discussion paper on generative AI-enabled medical devices, issued by the Digital Health Center of Excellence within CDRH.
- The proposed risk framework is two-axis: it separates a device's inherent generative-AI risk from the clinical risk of its intended use, rather than scoring risk on a single scale.
- Premarket evaluation is framed around "competency assessment" — combining non-clinical benchmarking of the model with clinical confirmation, a departure from the fixed-performance validation used for traditional locked-algorithm devices.
- The paper separately addresses foundation models and agentic AI systems — tools that act or make decisions with more autonomy than a static diagnostic algorithm — as a category needing distinct oversight.
- Public comments are due October 19, 2026 under docket FDA-2026-N-7874; the FDA is explicit that this paper does not represent final or proposed regulatory policy, only an early basis for discussion.
Reel Script
Hook (~15-20s): The FDA has never actually written rules for AI that behaves like ChatGPT inside a medical device. This week it published the first real attempt — and it splits risk into two separate questions instead of one.
Core Concept (~45-90s): Traditional FDA-cleared AI devices are "locked algorithms" — the same input always produces the same output, so regulators can validate them once and be done. Generative AI-enabled devices don't work that way: a model built on an LLM or foundation model can respond differently to the same prompt, adapt over time, or even act with some autonomy in what the FDA calls "agentic AI." The discussion paper proposes evaluating these devices on two separate axes — the inherent risk of the generative AI technology itself, and the clinical risk of the specific use case it's deployed for — rather than forcing both into one risk score the way locked-algorithm devices are today.
Hands-On (~45-150s): The paper lays out what premarket review could look like under this framework: "competency assessment," a combination of non-clinical benchmarking of the model's outputs plus clinical confirmation that it performs safely in the intended use. On the back end, it proposes risk-proportionate postmarket monitoring — higher-risk generative devices get watched more closely after launch, rather than every device getting the same fixed check-in schedule. The paper is explicit that foundation models and agentic AI systems need their own consideration separate from narrower generative tools, since an agentic system can take actions rather than just generate text or images. None of this is final: the FDA is collecting feedback from manufacturers, clinicians, and researchers under docket FDA-2026-N-7874 through October 19, 2026, before it writes anything binding.
Takeaway (~20-30s): If you build or deploy generative AI in a clinical product, this comment period is the moment to shape the rules before they're locked in — the two-axis framework and competency-assessment model are still soft clay, not settled policy. Health-tech teams should read the full paper and submit feedback before October 19.