The FDA Is Asking How It Should Regulate Generative AI

Artificial intelligence is already deeply embedded in medical technology. Generative AI, however, presents a different regulatory challenge.

On August 18, 2026, the FDA released a discussion paper exploring how generative AI-enabled medical devices should be evaluated and regulated. Rather than presenting a finished framework, the agency opened the issue for public comment through October 19, 2026, specifically inviting feedback from device manufacturers, clinicians, consumers, researchers, and other interested parties.

The timing is notable. An IntuitionLabs analysis of the FDA’s AI-Enabled Medical Device List reports a historical estimate of 1,451 authorized AI/ML devices by the end of 2025, with 295 added during 2025 alone. Approximately 76% of those listings were in radiology. IntuitionLabs cautions that these are third-party estimates based on the FDA’s non-comprehensive list rather than an official FDA count of every AI device authorization.

That existing ecosystem is largely built around more traditional AI and machine-learning applications. Generative AI introduces a new problem: systems capable of creating new content, interacting conversationally, relying on foundation models, and potentially operating with greater autonomy.

Why Generative AI Is Different

The FDA already regulates AI-enabled products when they meet the legal definition of a medical device. Many existing systems perform narrow tasks, such as analyzing medical images or physiological signals.

Generative systems can behave differently.

A model may produce text, images, recommendations, or other outputs in response to open-ended prompts. It may also rely on a foundation model developed by another company or operate as part of a more autonomous agentic system.

The FDA says these capabilities may introduce risks that differ from those associated with traditional software and AI-enabled medical devices. Its discussion paper therefore focuses on risk assessment, premarket evaluation, postmarket monitoring, foundation models, and agentic AI systems.

The American Hospital Association similarly characterized the paper as an effort to gather early stakeholder input before a broader policy framework is established.

How Do You Test Generative AI Before It Reaches Patients?

One possibility proposed by the FDA is a two-axis framework for risk assessment.

The agency is also considering a premarket evaluation model based on competency assessment, inspired at a high level by how physicians themselves are trained and evaluated. This could involve non-clinical benchmarking followed by clinical confirmation to determine whether a generative AI-enabled medical device performs as intended.

Postmarket oversight may be equally important.

Traditional medical software can be evaluated against relatively defined functions. Generative systems may behave differently depending on prompts, context, model updates, or changes to an underlying foundation model.

The FDA is therefore considering risk-proportionate postmarket monitoring rather than treating authorization as the end of the regulatory process.

The Existing AI Device Landscape Offers Some Clues

The FDA is not starting from zero when it comes to artificial intelligence.

According to IntuitionLabs, the number of AI/ML devices entering the market has risen rapidly, from only six in 2015 to an estimated 295 in 2025. Most have entered through the 510(k) pathway, and imaging remains by far the largest specialty.

But the tracker also highlights challenges that may become even more complicated with generative systems. Published analyses have identified gaps in reporting around study design, patient outcomes, training data, model characteristics, and postmarket performance.

Generative AI could make questions around transparency and lifecycle monitoring even more important.

Physicians Have an Opportunity to Weigh In

Perhaps the most interesting part of the FDA announcement is who is being asked to participate.

This is not a comment period directed only at AI companies and medical device manufacturers. The FDA explicitly lists clinicians among the stakeholders it wants to hear from.

That matters because physicians are increasingly encountering generative AI in real healthcare workflows.

How should accuracy be evaluated? When should clinicians verify AI-generated outputs? How should regulators respond when the underlying foundation model changes? What should happen when an AI system performs well in controlled testing but behaves differently after deployment?

Those are not solely engineering questions.

What Happens Next

The August discussion paper is not final guidance and does not itself establish new regulatory requirements. Instead, it represents an early stage in determining what a future regulatory framework could look like.

Generative AI is advancing quickly in healthcare. The FDA is now asking what evidence, oversight, and monitoring should accompany it when it becomes part of a regulated medical device.

For practicing clinicians, this is a relatively unusual opportunity to participate while those rules are still taking shape.

—

Sources: 

https://www.fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices

https://www.aha.org/news/headline/2026-08-18-fda-seeks-feedback-potential-regulatory-approaches-generative-ai-enabled-medical-devices

https://intuitionlabs.ai/articles/fda-ai-medical-device-tracker