The U.S. Food and Drug Administration is asking a consequential question: what should it take for a generative AI-enabled medical device to earn and keep trust?
On August 18, 2026, the agency released a discussion paper and opened a public feedback process covering risk assessment, premarket evaluation, postmarket monitoring, foundation models, and agentic AI systems. The FDA is not issuing new rules yet. It explicitly says the paper is for discussion and does not represent draft or final guidance. Still, the questions it asks are a useful early map for companies developing or buying health AI.
The shift is from static software to changing behavior
Traditional medical-device review often centers on a defined product performing a defined function. Generative systems can produce varied outputs, rely on foundation models, connect to external tools, and change through updates or configuration. That makes a simple one-time performance claim harder to sustain.
The FDA proposes exploring a two-axis approach to risk assessment and a premarket model built around competency assessment. In plain language, the agency is asking whether a system can demonstrate relevant capabilities in non-clinical testing and then confirm that performance in clinical settings before it reaches patients.
Why monitoring becomes part of the product
The paper also focuses on risk-proportionate postmarket monitoring. For health AI teams, that points to an operational reality: evidence cannot stop at launch. Teams may need to define what they will observe after deployment, how they will investigate performance concerns, and when a model, prompt, tool connection, or workflow change demands reevaluation.
This is especially relevant for products that summarize clinical information, support decisions, automate documentation, or coordinate actions across systems. The more a product can influence care, the more disciplined its boundaries, validation, and escalation paths need to be.
What builders and buyers can do now
Do not treat this paper as a compliance checklist. Treat it as a design review prompt. Inventory the clinical claim, intended user, model dependencies, failure modes, human review steps, and monitoring signals for each workflow. Separate a compelling demonstration from evidence that a device performs safely and effectively in its intended context.
The FDA is accepting comments under docket FDA-2026-N-7874 through October 19, 2026. Manufacturers, clinicians, researchers, patients, and other stakeholders can respond. The eventual policy may change, but the direction is clear: generative AI in medical devices will be judged not only by what it can generate, but by how reliably teams can evaluate and govern it.
The practical takeaway
For healthcare AI, the next differentiator may be less about a model's headline capability and more about the quality of the proof around it. Products designed with scoped use cases, documented evaluation, human accountability, and ongoing monitoring will be better prepared for both regulatory scrutiny and real clinical adoption.