On August 18, 2026, the FDA’s Digital Health Center of Excellence quietly published a discussion paper that could reshape how generative AI tools enter American hospitals. The document lays out the agency’s early thinking on regulating generative AI-enabled medical devices, from chatbot-style clinical assistants to AI systems that draft parts of a diagnosis. With the public comment window closing October 19, hospital groups, device makers and researchers are racing to weigh in before the FDA starts turning the paper into actual rules.
A Regulator Trying to Catch Up
Generative AI tools behave differently from the AI-enabled devices the FDA has cleared for years, which are mostly narrow classifiers: software that flags a suspicious mole or measures a heart rhythm from a single type of input. Generative systems can produce open-ended text, adapt their outputs based on context, and in some cases continue learning after deployment. The FDA’s discussion paper acknowledges that its existing premarket review framework, built around a fixed, validated device, was not designed for software whose behavior can shift. The agency is asking for input on how to assess risk, review performance before approval, and monitor these tools once they are in clinical use.
What Hospital Groups Are Pushing Back On
The American Hospital Association submitted formal feedback through a letter sent to the FDA, urging the agency to be explicit about which generative AI use cases actually need premarket clearance versus which can be treated more like clinical decision-support software that doctors retain full discretion over. Hospital systems have told the agency they are wary of a framework that is so broad it captures low-risk documentation tools alongside high-risk diagnostic assistants, warning that overly cautious rules could slow adoption of tools hospitals already rely on for administrative relief, such as ambient scribes, without meaningfully improving patient safety.
The Backdrop: A List That Already Has 1,600 Entries
The debate is not happening in a vacuum. The FDA’s running list of AI-enabled medical devices had grown to roughly 1,600 entries by its September 2026 update, though the agency itself cautions that the list is not a comprehensive accounting of every AI tool in clinical use, since many fall outside the agency’s clearance requirements entirely. That caveat is part of what is fueling the current debate: a large and growing share of AI software touching patient care may never pass through FDA review at all, which is exactly the gap this discussion paper is trying to address for the generative AI subset.
A Separate Deregulatory Move Complicates the Picture
Adding to the uncertainty, the FDA said in January 2026 that it would loosen oversight of certain clinical decision support software, including some AI tools that assist with diagnosis, allowing them to reach the market without full FDA review as long as they meet the agency’s existing criteria for exemption. Critics argue the generative AI discussion paper and the clinical decision support carve-out pull in opposite directions — one tightening scrutiny on generative tools, the other loosening it for decision-support software broadly — and are asking the FDA to reconcile the two before finalizing any new framework.
Postmarket Monitoring Is the Harder Problem
Premarket review has dominated the public conversation, but agency officials and outside commenters agree the tougher question is what happens after a generative AI tool is cleared. A separate FDA request for comment in 2025 asked how to measure AI device performance once deployed in real clinical settings, where patient populations, documentation habits and even hospital workflows vary widely from the conditions under which a device was validated. For generative tools capable of producing different outputs from the same input, commenters argue postmarket surveillance will need to catch performance drift in close to real time, a capability few hospitals currently have the infrastructure to support.
What Comes After October 19
Once the comment window closes, the FDA is expected to use the feedback to inform a future draft guidance document, a process that historically has taken months rather than weeks. For hospital IT departments and AI vendors alike, the stakes are less about any single rule than about predictability: health systems say they need clarity on which generative AI products will require clearance before they can commit budgets to large-scale deployments. Whatever framework the agency eventually lands on will likely become the template other countries look to as they write their own rules for AI in medicine, making this quiet comment period more consequential than its low public profile suggests.
Photo: cottonbro studio / PEXELS via Pexels
