Frontier AI systems are becoming more powerful by the month, and at a pace that seems to be accelerating. However, from the standpoint of U.S. economic growth and competitiveness, our technical lead in AI only really matters to the extent that AI can diffuse into tangible, real-world applications. Teams of AI agents can now build entire software projects without human intervention, but only because software development is relatively unregulated. Translating recent AI progress from the realm of bits to the realm of atoms is usually much harder, in large part because of the litany of legal and regulatory frameworks designed for the pre-AI age.
Healthcare is perhaps the clearest sector where AI can deliver substantial near-term benefits to ordinary Americans, provided that outdated regulations don’t get in the way. Consider that multiple medical AIs now score 100 percent on the U.S. Medical Licensing Exam and have become increasingly reliable at assigning medical diagnoses—even more reliable than doctors in controlled experiments.
To ensure the safety of AI-powered health and medical applications while still reaping their benefits, governments will need to modify many laws. But since legislative bodies can be slow to change statutes, it is often easier to experiment with new governance structures implemented wholesale—for example, through a regulatory sandbox. Utah’s recent experiment in regulatory sandboxing offers a useful case study.
In March 2024, Utah enacted the AI Policy Act into law. Among other measures, it contained a sandboxing provision: the Office of AI Policy was established with the authority to grant temporary waivers of specific laws that might inhibit the use of AI technologies through “regulatory mitigation agreements.”
Enter Doctronic, a company developing an AI tool to review prescription refill requests. Ordinarily, Utah law requires a doctor to issue prescriptions. However, under the sandboxing program, Doctronic signed a mitigation agreement with the Office of AI Policy that exempted the company from enforcement of professional licensing rules for one year, allowing Doctronic’s AI tool to make prescription renewal decisions for a list of 192 low-risk drugs. Patients still need to visit a doctor for new prescriptions but can now refill their existing prescriptions using AI. The agreement waives no other laws and, importantly, emphasizes that health data privacy and cybersecurity laws still apply.
Utah’s regulatory sandbox program offers lessons for experimenting with new governance structures more generally. Rather than creating ad hoc exemptions for novel AI applications in statute, the Utah legislature delegated broad waiver authorities to a new agency with a mandate to experiment with new regulatory models, thereby side-stepping incumbent interest groups who can more easily lobby against reforms on a case-by-case basis.
Utah’s framework empowers the Office to grant mitigation agreements to specific companies. In addition to raising risks of preferential treatment, this approach poses challenges for scalability. While an agency can act faster than a legislature, there is still a good deal of paperwork involved in preparing a plan for the regulatory waiver. This is a reasonable approach for piloting new AI applications and collecting data needed to inform more permanent reforms, but it is not a long-term solution on its own. As early data come in, regulatory sandboxes should transition from granting company-based waivers to waivers pegged to performance-based benchmarks—benchmarks that could then help inform legislatures on the shape of alternative governance mechanisms. This presents greater upfront challenges in defining what standards and evaluations an AI application provider must meet in any given use case, but has the potential to allow for a more streamlined approval process going forward.
More far-reaching governance ideas such as special economic zones or Freedom Cities could improve scalability while facing other tradeoffs. By designating geographic areas in which a wide cross-section of legacy regulations are waived, governments can test new laws and facilitate industrial or technological development. Just as states are the “laboratories of democracy,” special economic zones have the potential to allow for even more fine-grained iteration on sets of laws. SEZs are not without risk, as the difference in rules can create opportunities for regulatory arbitrage and a race to the bottom. Nevertheless, as technological progress forces governmental adaptation, it will also necessitate experimentation. On this front, Utah’s AI Policy Act is leading the way.
