Key Takeaways
- Licensing autonomous clinical AI—modeled after physician licensure—offers a flexible, competency‑based alternative to the FDA’s current Software as a Medical Device (SaMD) framework.
- The proposal calls for AI models to pass standardized medical licensing exams (USMLE steps) and relevant specialty board tests before any deployment.
- After certification, AI would undergo a supervised, residency‑like deployment period to demonstrate non‑inferior clinical performance.
- Licensure would define a clear scope of practice, be time‑limited, and require periodic reevaluation to account for AI’s adaptive nature.
- Responsibility would be layered: developers answer for model performance; deploying institutions handle integration, monitoring, and adverse‑event reporting.
- Federal pre‑emption would prevent a patchwork of state‑level licenses while preserving state authority over scope, supervision, and enforcement.
- The authors acknowledge that licensure does not resolve broader issues such as data governance or workforce impact, which may need separate legislation.
The Regulatory Gap Facing Autonomous Clinical AI
Rapidly advancing autonomous clinical AI presents a regulatory challenge that the FDA’s current framework, built for drugs and devices, poorly addresses. “That framework assumes static products, narrow indications, and clear manufacturer accountability, assumptions that autonomous clinical AI defies,” write Alon Bergman, PhD, Robert Wachter, MD, and Ezekiel “Zeke” Emanuel, MD, PhD, in their JAMA opinion piece. The trio argue that treating AI as a mutable, learning system demands oversight more akin to how physicians are licensed and monitored.
Why Physician‑Like Licensure Fits AI
Licensure, they add, is “better aligned with the realities of autonomous clinical AI.” Unlike a fixed device, an AI model can improve—or deteriorate—over time as it ingests new data. A licensing approach would evaluate competency upfront, then track performance continuously, mirroring the lifelong learning and re‑certification cycles required of human clinicians.
Competency Certification Through Standardized Examinations
The first step in their proposal is competency certification via standardized examinations. Bergman and colleagues envision a future in which every autonomous AI clinical model is tested on all three components of the U.S. medical licensing examination (USMLE): scientific principles, clinical knowledge, and preparedness for decision‑making. Before being greenlit by the FDA, these models would have to beat or equal the median scores of recent human exam‑passers. Each would also have to prove accurate and reliable on any relevant specialty board examinations aligned with the model’s intended scope. “Passing scores would establish minimum competency, not clinical readiness,” the authors write, emphasizing that the exam is a floor, not a guarantee of expert judgment.
Supervised Clinical Deployment Resembling Residency
Models that meet the threshold would enter a supervised clinical deployment period. Such a period would emulate residency training, as models would have to demonstrate non‑inferior clinical performance. “Specific requirements for patient volume would be established through the multistakeholder standards process,” the authors explain. During this phase, AI would operate under direct clinician oversight, allowing real‑world assessment of diagnostic accuracy, treatment suggestions, and interaction with electronic health records while patient safety remains safeguarded by supervising physicians.
Defining a Scope of Practice
This licensure‑like approach would define a scope of practice. In so doing, the process would specify which clinical functions specific AIs may perform, in which settings—and under what level of oversight. “An AI certified for primary care triage,” the authors offer as an example, “could gather histories and recommend next steps but could not prescribe care or medications without clinician approval.” By delineating permissible actions, the framework prevents overreach while still enabling AI to handle routine, high‑volume tasks such as screening, risk stratification, or preliminary documentation.
Time‑Limited Certification and Ongoing Reevaluation
Certification would be time‑limited. It would also be contingent on ongoing performance monitoring and reevaluation, Bergman and colleagues propose. “Because AI systems are adaptive rather than static, continued authorization would depend on periodic (eg, biennial) demonstration of acceptable levels of clinical performance.” This mirrors the maintenance of certification (MOC) requirements for physicians, ensuring that any drift in model behavior—due to data shift, algorithm updates, or emergent biases—is caught early and corrected before patient harm occurs.
Explicit Accountability Structures
Accountability structures would be explicit. AI developers would bear primary responsibility for model performance, the authors maintain. Performance measures would mandatorily include safety, accuracy and behavior across clinical contexts. Meanwhile, deploying institutions would be responsible for implementation, the team suggests. Aspects of implementation would include integrating the AI into clinical workflows, monitoring patient outcomes and reporting adverse events, and ensuring that the conditions covered under the AI approval are followed—including triaging to a clinician when appropriate. “This layered structure mirrors existing liability frameworks in medicine,” the authors point out, aligning AI responsibility with the shared‑blame model already used for drugs, devices, and practitioner error.
Federal Pre‑emption to Avoid a State‑by‑State Patchwork
Federal pre‑emption would be necessary to avoid 50 parallel licensing regimens. “Under this approach, federal certification of AI competency would be binding on all states, whereas states would retain authority over scope of practice, supervision requirements and enforcement,” Bergman, Wachter and Emanuel write. A national baseline would prevent manufacturers from navigating contradictory state rules, while still allowing localities to tailor supervision levels—such as requiring more intensive oversight in rural settings or for high‑risk specialties.
Advantages Over the Current FDA SaMD Framework
More: “Recognizing autonomous AI’s rapid evolution, this licensure‑like oversight would subject autonomous clinical AI models to more continuous, comprehensive, flexible and clinical evaluation than the FDA’s current Software as a Medical Device framework.” The SaMD approach treats AI as a static product with predefined indications, requiring new submissions for each algorithm update. By contrast, licensure embraces the iterative nature of machine learning, allowing updates to be evaluated through ongoing performance metrics rather than repeated premarket reviews.
Limitations and the Need for Further Regulation
The scholars acknowledge that licensing AI would not settle questions around overarching concerns such as data governance and workforce transformation. These and other matters, they suggest, probably call for new regulation or legislation. Issues like patient consent for data use, algorithmic transparency, equity across diverse populations, and the potential displacement or retraining of clinical staff remain outside the scope of a licensure model. Nevertheless, they argue that as clinical AI “increasingly resembles clinicians in its capabilities, our regulatory frameworks must evolve accordingly.”
By anchoring oversight in physician‑like licensure, the proposal offers a pragmatic, adaptable pathway to harness AI’s benefits while preserving the safety and accountability that patients and providers expect.
https://healthexec.com/topics/artificial-intelligence/licensing-clinical-ai-its-doctor-3-medical-leaders-issue-6-step-call