Three Medical Leaders Propose Licensing Framework for Clinical AI
Three prominent researchers are calling for clinical AI systems to be licensed like physicians-a regulatory approach they argue would better oversee the technology than current FDA rules while preserving its capabilities.
Alon Bergman, PhD, Robert Wachter, MD, and Ezekiel Emanuel, MD, PhD, laid out their case in an opinion piece published by the Journal of the American Medical Association. They contend that the FDA's existing framework, designed for drugs and devices, assumes static products with narrow uses and clear manufacturer accountability. Autonomous clinical AI defies all three assumptions.
"As clinical AI increasingly resembles clinicians in its capabilities, our regulatory frameworks must evolve accordingly," the authors write.
Licensure would represent a step forward from the FDA's current Software as a Medical Device framework, which subjects AI to less continuous evaluation than the authors' proposal would require.
The Six-Step Proposal
1. Competency certification through standardized exams. Each autonomous AI model would be tested on all three components of the U.S. medical licensing examination: scientific principles, clinical knowledge, and decision-making preparedness. Models would need to match or exceed the median scores of recent human exam-passers, plus perform accurately on relevant specialty board examinations tied to their intended scope.
Passing scores would establish minimum competency, not clinical readiness, the authors note.
2. Supervised clinical deployment. Models meeting the threshold would enter a supervised period modeled on residency training. They would have to demonstrate clinical performance equal to or better than existing standards, with specific patient volume requirements determined through a multistakeholder process.
3. Defined scope of practice. Licensure would specify which clinical functions an AI can perform, in which settings, and under what level of oversight. An AI certified for primary care triage could gather patient histories and recommend next steps but could not prescribe medications without clinician approval.
4. Time-limited certification with ongoing monitoring. Continued authorization would depend on periodic-perhaps biennial-demonstration of acceptable clinical performance. Because AI systems adapt over time, static approval would not suffice.
5. Clear accountability structures. AI developers would bear primary responsibility for model performance, measured by safety, accuracy, and behavior across clinical contexts. Deploying institutions would be responsible for integration into clinical workflows, patient outcome monitoring, adverse event reporting, and ensuring conditions of approval are met.
This layered structure mirrors existing liability frameworks in medicine, the authors say.
6. Federal preemption. Federal certification of AI competency would be binding on all states, though states would retain authority over scope of practice, supervision requirements, and enforcement. This prevents the creation of 50 parallel licensing regimens.
What the Proposal Does Not Address
The authors acknowledge that licensure would not resolve broader concerns around data governance and workforce transformation. Those issues likely require separate regulation or legislation.
The proposal focuses specifically on how individual AI models can be certified to practice, not on the systemic questions surrounding how clinical AI reshapes the medical workforce or how patient data flows through these systems.
Learn more about AI for Healthcare and Generative AI and LLM systems that power autonomous clinical tools.
Your membership also unlocks:
