Public comments on the FDA’s draft principles for generative AI in healthcare closed recently, and the themes running through the responses are consistent. Hospitals and device makers want clarity. Patient groups want accountability. Nearly everyone wants to know who is responsible when a generative system produces a confident, wrong clinical recommendation.
The draft framework, developed with the European Medicines Agency, is deliberately high-level. It sets expectations around transparency, risk management and lifecycle monitoring without prescribing specific technical controls.
The deployment question
Most of the difficult cases involve generative AI used as a decision-support tool inside a clinical workflow: summarizing records, drafting notes, suggesting differential diagnoses. These uses sit in a gray zone between a regulated medical device and general productivity software.
Hospitals are already deploying such tools under existing governance processes, often faster than their internal policies anticipated. The FDA’s principles signal that “we bought it from a vendor” will not be a sufficient answer if something goes wrong.
Lifecycle monitoring is the hard requirement
A model that performed acceptably at deployment can degrade as clinical practice changes, as patient populations shift, or as it is used in ways the developer did not test. The draft principles emphasize ongoing monitoring, which implies logging, auditing and the organizational capacity to act on what the logs show.
Many health systems do not currently have that capacity. Building it is an operational cost that procurement decisions rarely include.
What comes next
Draft principles become guidance, and guidance eventually shapes enforcement. The practical effect over the next year will be felt in contracting: hospitals asking vendors for monitoring commitments, audit rights and clearer statements of intended use. The regulatory framework is still forming, but the liability conversation has already moved.