Generative AI guidance is coming from the U.S. Food and Drug Administration as the agency works to clarify how rapidly evolving artificial intelligence should be regulated when incorporated into medical devices. Rick Abramson, director of the FDA’s Digital Health Center of Excellence, said the agency’s objective is to develop formal policy guidance covering generative AI broadly, along with narrower guidance addressing particularly complex applications. The comments signal a transition from exploratory discussions toward a more defined regulatory framework for healthcare AI.
The announcement follows an FDA discussion paper released on August 18, 2026, examining risk assessment, premarket evaluation and postmarket monitoring for generative AI-enabled medical devices. The agency is accepting public feedback through October 19, 2026.
Generative AI Guidance Moves Toward Policy
Abramson told STAT that the FDA ultimately intends to publish documents explaining its policies for generative AI-enabled medical devices. He indicated that developers should expect both broad guidance and specialized guidance for topics involving unusual complexity or regulatory interest. No specific publication timeline has been announced.
The FDA’s current discussion paper is therefore not itself draft or final guidance. The agency explicitly states that the document is intended to gather stakeholder feedback and does not establish new regulatory expectations or policies.
That distinction matters for medical-device developers. The concepts being discussed may influence future policy, but companies should not treat them as finalized requirements.
Generative AI Guidance Addresses New Risks
Traditional medical software often operates with relatively predictable inputs and outputs. Generative AI is fundamentally different because the same system can accept diverse prompts and create open-ended responses.
This makes exhaustive testing difficult.
The FDA notes that the number of possible inputs and outputs produced by generative systems can be too large for traditional software-testing methods to cover comprehensively. That challenge becomes particularly significant when AI is used for clinical communication, decision support, patient triage or other healthcare functions.
Regulators therefore need methods capable of evaluating whether a system remains reliable across a broad range of realistic clinical situations.
FDA Considers Competency-Based AI Testing
One of the most notable ideas in the FDA discussion paper is a competency-based approach to premarket evaluation.
The concept draws inspiration, at a high level, from how physicians are assessed and credentialed. Rather than attempting to test every possible interaction individually, regulators could evaluate whether an AI system demonstrates appropriate competency for the clinical tasks it is designed to perform.
The FDA describes a potential framework involving two major elements: nonclinical device benchmarking followed by clinical confirmation that the system performs as intended.
Generative AI Guidance Could Focus on Competency
Competency testing could provide a different way to evaluate systems that produce highly variable outputs.
For example, a generative AI tool designed to support clinical triage might be evaluated on whether it consistently recognizes situations requiring escalation rather than whether it reproduces a fixed answer.
The FDA is also considering both under-escalation and over-escalation risks. Missing an urgent condition could create obvious patient-safety concerns, while excessive escalation could overwhelm healthcare resources or create unnecessary treatment.
The acceptable balance may differ according to the clinical context and intended use.
Risk Could Determine FDA Requirements
The FDA discussion paper outlines a potential two-axis framework for assessing generative AI medical-device risk.
Regulatory expectations could vary depending on the type of function performed and the potential consequences when a system produces an incorrect result.
A tool performing a relatively low-risk administrative function would therefore present different regulatory concerns from an AI system involved in diagnosis, treatment recommendations or clinical escalation.
This risk-proportionate approach could help avoid applying identical requirements to every generative AI application.
At the same time, developers would need to define intended uses carefully and demonstrate appropriate safeguards for the risks associated with each function.
Postmarket Monitoring Gains Importance
Generative AI also raises questions that continue after a device reaches the market.
Traditional regulatory review largely evaluates evidence produced before commercialization. However, generative systems may behave differently when exposed to changing patient populations, new workflows or novel inputs.
The FDA is therefore considering risk-proportionate postmarket monitoring approaches.
Ongoing monitoring could help identify performance changes, unexpected behaviors and emerging safety concerns that were difficult to predict during premarket testing.
This could become especially important if regulators accept greater uncertainty before market entry in exchange for stronger real-world surveillance afterward. However, the agency is still collecting feedback on what such monitoring should involve.
Foundation Models Create Regulatory Complexity
The discussion paper also specifically addresses foundation models and agentic AI systems.
Foundation models can support multiple downstream applications rather than one narrow healthcare task. A single model could potentially contribute to documentation, clinical communication, decision support and other functions.
Agentic systems create another challenge because they may perform sequences of actions rather than simply generate one response.
Generative AI Guidance Must Address Intended Use
These capabilities complicate conventional concepts of intended use.
Regulators may need to determine whether evaluation should focus primarily on the underlying model, the specific medical-device application built around it or the complete workflow in which the technology operates.
The FDA has not announced a final answer. Its discussion paper asks stakeholders to help define approaches that remain scientifically rigorous while accommodating the unusual capabilities of generative AI.
Developers Should Prepare for Greater Scrutiny
Although final policy has not been released, the FDA’s direction provides useful signals for healthcare AI developers.
Companies should expect increasing attention to:
- Clearly defined intended use
- Clinical risk and potential patient harm
- Representative performance testing
- Competency across realistic scenarios
- Human oversight
- Data quality and governance
- Postmarket performance monitoring
- Foundation-model dependencies
- Updates and performance changes
- Documentation supporting regulatory submissions
Developers should also avoid assuming that conventional AI validation methods will automatically satisfy requirements for generative systems.
The FDA’s existing January 2025 draft guidance for AI used in drug and biological product regulatory decision-making already emphasizes risk-based credibility assessment tied to a clearly defined context of use.
Generative AI Guidance Signals Regulatory Shift
The FDA’s latest activity demonstrates that generative AI is moving from regulatory uncertainty toward more formal oversight.
Abramson’s commitment to future broad and specialty guidance suggests the agency recognizes that one regulatory document may not adequately address every use of the technology.
For developers, healthcare organizations and investors, clearer policy could reduce uncertainty about the evidence required before an AI-enabled medical device reaches patients.
For regulators, the challenge is more complex. They must establish standards that protect patients without freezing innovation in a technology category that can evolve quickly.
The current FDA discussion paper represents an early step in that process. Public feedback will help shape how regulators approach competency testing, risk classification, foundation models and postmarket monitoring.
Ultimately, generative AI guidance could establish important boundaries for how increasingly capable AI systems are developed, evaluated and monitored across healthcare.
For more healthcare industry updates, insights and news, visit DistilINFO. Click here to subscribe to stay informed.
