Generative AI is already helping clinicians interpret test results, administer antibiotics, and respond to patient questions. The current health AI approval process uses the Food and Drug Administration’s (FDA) Software as a Medical Device (SaMD) framework, which is ideal for single-manufacturer tools designed to perform narrow tasks, such as classifying high-risk lung lesions. But modern generative AI products have a broader skillset and can be built on the foundation created by one company, tweaked by another, and extended via third-party plug-ins, without clear regulatory responsibilities.
In a recent paper, Penn LDI Senior Fellow Eric Bressman and co-authors propose: Japan Automobile Manufacturers Association The perspective part, a different approach: License AI in much the same way you license human doctors, nurses, and physician assistants to work together under supervision.
Bresman and his coauthors argue that concerns about generative AI, such as hallucinations and performance fluctuations, reflect late-19th-century concerns about quack treatments and variable clinician training. A licensing approach that combines standards of practice with ongoing monitoring and education can be adapted to AI regulation.
Ideally, a new federal Digital Licensing Commission would oversee this framework, Bressman and colleagues suggest.
However, existing federal and state agencies could play an important role. The FDA may retain its role in premarket evaluation, eliminating the need for developers to submit to licensing authorities in 50 states.
A designated health system with AI expertise could serve as an implementation center, with ongoing oversight and discipline left to state medical boards and a federal coordinating agency to harmonize standards.
“A licensing framework may help ensure that innovation scales responsibly, rather than pre-empting liability,” the authors write. The table below shows similarities between clinician licensing and possible future AI licensing structures.
| License concept | human clinician | Generation AI |
| Pre-licensing requirements | Must have an accredited degree and passed a national certification exam.
Clinical training guidance period |
Technical validation of predefined competencies (AI National Board Examination)
Supervised pilot at a nationally certified “Implementation Center” (AI “Resident”) |
| Scope of implementation | Overview of approved medical services (which residents can use independently)
PA/NP cooperation or supervision agreement |
Delineation of authorized functions (e.g. image interpretation) (groups with degree of autonomy)
Guidance on supervising clinician supervision for each function |
| Institutional Credentials | Health system credentials to perform specific steps and review results. Permissions can be suspended if there are safety concerns | The health system’s AI governance committee will scrutinize site-specific implementations, determine local authority within the scope of authorized AI practice, and monitor local quality metrics. You can revoke permissions or deactivate the model if thresholds are not met. |
| continuous monitoring | Continuing education requirements and periodic knowledge assessments (to maintain board certification) | Rerun updated benchmarks annually for each competency and report clinical performance measures for board review. |
| discipline/responsibility | State medical boards can investigate complaints and issue fines, suspensions, revocation of licenses, or mandatory retraining.
Conduct reported to the National Practitioner Databank Manufacturers, health systems, and clinicians may be held responsible |
The Digital Commission receives and processes complaints and can place AI systems on probation. Requires model patches or additional guardrails. Suspend or revoke your license
Maintain a public database of disciplined models and corrective action plans For restricted licenses or high-risk functions, the responsibility lies with the supervising clinician and facility. Developers are responsible for higher degrees of autonomy (e.g., low-risk features) |
Published in the November 17, 2025 issue of “Software as a doctor — is it time to license artificial intelligence?” JAMA Internal Medicine. Authors include Eric Bressman, Carmel Shachar, Ariel D. Stern, Ateev Mehrotra, and more.
author
Learn more about AI in healthcare
blog post
AI in healthcare
Getting a second opinion: Generative AI generates inconsistent medical recommendations
Study of six large-scale language models reveals significant differences in responses to clinical scenarios
AI in healthcare
AI is ushering medical school into a new era, but are you ready?
Penn LDI Panel Experts Call for Rapid Training of Students and Faculty
news
AI in healthcare
National Academy of Medicine issues code of conduct to guide AI revolution in healthcare
Penn’s Kevin B. Johnson, one of the authors, explains the principles set forth in this book.
news
AI in healthcare
Penn CHIBE withdrawal focuses on new AI tools to provide evidence to policymakers
Envisioning a chatbot with a more focused and comprehensive large-scale language model
news
AI in healthcare
In-the-loop or on-the-loop: The challenge of AI clinical decision support
2025 Penn Nudge in Healthcare Symposium will focus on human-machine interfaces
blog post
AI in healthcare
AI breakthroughs could transform public health campaigns
This social media tool to fight HIV could be scaled up to help local health authorities, says LDI expert
