Our Take
The proposal arrives after the regulatory failure, not before it, which tells you everything about academic AI policy work.
Why it matters
Utah is one of 47 states writing 250+ clinical AI bills with zero coordination. The FDA device framework can't handle adaptive models that update continuously.
Do this week
Healthcare AI teams: Document your competency benchmarks now before federal standards lock in requirements you can't meet retrospectively.
Utah suspended AI prescription program after licensing board revolt
Utah's Medical Licensing Board called for immediate suspension of the state's Doctronic pilot program in late April. The program allows AI to evaluate patients and recommend prescription renewals for nearly 200 chronic condition drugs. Utah planned to phase out physician review of each case once safety benchmarks were met.
The board learned about the pilot only after launch and warned that proceeding without clinical oversight "potentially places Utah citizens at risk." The controversy centers on sequence: Utah deployed first, then added oversight, rather than proving competency before independent practice.
Penn medical ethics professor Alon Bergman and colleagues published a federal licensing framework in JAMA on April 29, calling for a new Office of Clinical AI Oversight within HHS. The proposal requires AI systems to pass USMLE exams, complete supervised deployment phases, and maintain defined scopes of practice with biennial renewal.
State fragmentation blocks deployment while physician shortages worsen
Forty-seven states are considering over 250 clinical AI bills, creating regulatory patchwork. California bars AI from denying insurance coverage. Colorado mandates bias assessments. Each adds compliance costs without resolving core competency questions.
The physician shortage compounds urgency. National projections show tens of thousands of doctor shortfalls over the next decade (per the article's cited projections), especially in primary care and rural areas. Traditional fixes like more medical school seats take years to impact supply.
Recent AI performance data supports the case for autonomous systems. A 2025 study of nearly 40,000 primary care visits in Kenya found AI-supported clinicians made substantially fewer diagnostic errors (per the published study). The December NOHARM trial showed doctors did not outperform leading language models on any measured clinical task dimension.
Federal preemption would standardize competency, preserve state enforcement
The proposed framework includes four elements: demonstrated competency through USMLE performance and supervised deployment, defined scope of practice with escalation protocols, ongoing monitoring with time-limited renewal, and federal oversight with layered accountability.
Developers would bear primary responsibility for model performance. Deploying institutions would handle workflow integration and adverse event reporting. States would retain authority over scope of practice and enforcement but could not impose duplicate competency assessments.
Implementation would require congressional authorization to transfer autonomous clinical AI authority from FDA to the new HHS office. Developer user fees, modeled on FDA's device review funding since 1992, would build evaluation infrastructure.
The proposal addresses three main objections: state control concerns by noting AI crosses state lines instantly like telemedicine, equivalence worries by maintaining clinician roles for complex judgment, and implementation capacity through fee-funded infrastructure development.