FDA Opens Door to Clinician-Style Tests for Medical AI

Generative AI medical devices may eventually be judged less by whether regulators can anticipate every possible prompt and more by whether a finished product can reliably carry out its intended clinical task. The FDA has opened a public discussion on that possibility, including clinician-style competency testing, clinical confirmation, and monitoring after deployment.

The agency released Considerations for the Regulation of Generative AI-Enabled Medical Devices on August 18. It is seeking feedback through October 19, 2026, but the paper is neither draft nor final guidance. That distinction matters: the document starts a regulatory-design process rather than imposing new requirements on developers or healthcare providers.

The problem is that generative systems do not behave like conventional medical software with a defined function and a predictable range of inputs and outputs. They can take open-ended instructions, return different answers to similar prompts, and change when their underlying model, safeguards, or data sources change. Testing every possible interaction may therefore be impractical.

The FDA is instead considering an approach in which a manufacturer demonstrates that the completed device competently performs its intended clinical task. The assessment would focus on the product patients and clinicians actually use, rather than evaluating a foundation model in isolation. One general-purpose model can support products with different prompts, interfaces, safeguards, and clinical uses, making that distinction consequential.

A possible evidence path has three parts

• Nonclinical benchmarking could examine clinical knowledge, analytical ability, safety behavior, communication, and performance across patient groups and operating conditions.
• Clinical confirmation would test whether the device works with patients, clinicians, and real workflows. The needed evidence would vary with intended use and the potential harm of an incorrect output; a prospective clinical trial would not necessarily be required in every case.
• Post-deployment oversight could include periodic retesting, clinician review of real-world outputs, and monitoring for performance deterioration. Relevant software or third-party foundation-model changes could trigger another assessment.

The paper also reaches beyond systems that generate text or other outputs in a single exchange. It covers foundation models and agentic systems that can plan and execute multistep tasks, alongside premarket testing and risk-proportionate postmarket monitoring. The effort is led by the FDA’s Digital Health Center of Excellence within the Center for Devices and Radiological Health.

The clinician analogy gives the proposal a practical frame: regulators can test underlying knowledge, observe performance in clinical settings, and maintain oversight without predicting every situation a physician will face. The FDA is examining whether a version of that model can work for AI while accounting for the technical and legal differences of regulated devices.

The unresolved work is substantial. The paper identifies possible approaches and asks targeted questions intended to shape a future framework; it does not settle what evidence will be sufficient for a given device, when reassessment will be required, or which proposals will become policy. Stakeholders now have until October 19 to contest, refine, or support the model before the FDA decides its next step.