This month, the SANDBOX Act was introduced in the US Senate, marking the first attempt to establish a federal artificial intelligence regulatory sandbox. For the unfamiliar, a regulatory sandbox is a legal framework that allows organizations to experiment with emerging technologies in a controlled environment—providing regulatory enforcement relief in return for agency oversight. The aim is to create regulatory space for innovations to be tested while generating evidence to guide future reform.
Under the SANDBOX Act, the director of the Office of Science and Technology Policy (OSTP) would be charged with coordinating a cross-agency AI sandbox. Companies could apply for waivers from regulations they identify for two years, renewable up to four times. In exchange, firms must provide periodic progress and safety reports, post-consumer-facing risk disclosures, maintain and make available to the OSTP director any internal documentation and data, and report serious incidents of harm within 72 hours.
Like many early legislative efforts, this is a promising yet imperfect framework, presenting both benefits and potentially serious risks that must be addressed. That said, the goal of a federal AI sandbox is worth pursuing, as is the effort to refine and strengthen the design of this basic approach. To understand why, I’ll first examine the benefits a well-designed federal AI sandbox could deliver—including innovation acceleration and enhanced crisis management capabilities—before addressing framework risks that must be mitigated to fully realize these benefits.
Innovation Acceleration
The greatest advantage of regulatory sandboxes is their ability to enable the testing and deployment of new innovations. Regulation typically lags technological change, and outdated requirements can slow or prevent the development of new tools like AI systems. Sandboxes offer a policy correction. Through temporary regulatory relief, companies gain access to the testing and real-world refinement needed for innovation.
A growing body of global evidence suggests these programs can indeed deliver these positives. According to a comprehensive 2020 World Bank analysis of global sandbox programs, 88 percent of regulators reported that sandboxes successfully attracted new innovators and experimentation to their markets. More importantly, 55 percent noted these programs helped lower barriers to entry—a critical outcome for under-resourced innovators—while fostering the market competition essential for progress.
While regulatory sandboxes remain a relatively new policy concept—the oldest domestic program dates back just seven years—tangible outcomes back these survey results. Arizona’s fintech sandbox, for example, enabled startup Velex to test AI-powered cash-flow analysis as a credit assessment alternative to traditional credit scores. Such models have the potential to expand financial access to consumers with limited credit histories. Utah’s legal services sandbox, meanwhile, allowed startup Rasa to pilot software to help automate typically costly and time-consuming criminal record expungement.
Beyond this direct market facilitation, sandboxes offer a second critical innovation benefit: regulatory reform. These programs function not merely as tools of regulatory flexibility but also as policy experimentation laboratories. By providing carefully monitored regulatory relief, they enable agencies to test the effectiveness of existing rules, observe how regulations shape emerging technologies, and collect evidence to guide future policy decisions. World Bank findings underscore this experimentation value: among global agencies operating sandbox programs, a remarkable 50 percent report revising regulations based on evidence gathered through these initiatives. In fast-moving fields like AI, sandbox-driven reform may prove essential for pacing regulations with innovation, enabling the sustainable progress that only up-to-date policy can deliver.
For AI development, this track record of both innovation and reform success demands attention. While regulations have not significantly hindered consumer-grade AI innovation, the same cannot be said for deployments in highly regulated sectors, including healthcare, transportation, and finance. Recent transportation sector policy history illustrates the potential benefits of AI sandboxing. Today’s autonomous vehicle technology benefited, and perhaps even exists, precisely because regulators created sandbox-like programs such as California’s Autonomous Vehicle Tester (AVT) Driverless Program, which allowed monitored experimentation with then-prohibited driverless technology. By allowing experimentation, authorities directly enabled this life-saving innovation to be developed.
Under the broad flexibility of a federal AI sandbox, we can expect similar breakthroughs across other regulated domains. In the process, regulators can gather evidence to guide the long-term reforms necessary to foster sustained innovation.
Enhanced Crisis Management
Beyond innovation, sandboxes offer a less discussed but equally important advantage: crisis-ready regulatory flexibility. When societal disruptions strike—whether pandemics, cyberattacks, or other unforeseen emergencies—effective solutions often require adjusting regulation so research, testing, and deployment of technological solutions can proceed without delay. Permanent sandbox programs for AI and other essential technologies provide exactly this capability: a ready-made mechanism for companies and regulators to identify rules that can be briefly relaxed so technology can meet urgent societal needs.
While still speculative in the AI context, other crisis-ready sandboxes have already demonstrated value during past emergencies. During the COVID-19 pandemic, the United Kingdom—which had already established sandbox programs across multiple sectors—leveraged them effectively for crisis management. The Financial Conduct Authority’s fintech sandbox incentivized innovators to develop tools for crisis financing and COVID-19-related fraud detection. Meanwhile, the Information Commissioner’s Office used its privacy sandbox to support public health IT projects, including a permission-to-contact service that helped rapidly recruit volunteers for vaccine trials.
These UK examples pale in comparison to an even more consequential domestic tool: the Food and Drug Administration’s Emergency Use Authorization (EUA) authority. Arguably a regulatory sandbox, this mechanism allowed COVID-19 vaccines to be distributed on a heavily monitored basis to meet emergency needs. Precisely because US regulators had this tool on hand to temporarily relax drug approval processes, we were able to develop and distribute the solution that ended the crisis.
In AI policy, concerns over such pandemic-scale safety crises drive much of the mainstream policy debate. Despite the prevalence of these important concerns, however, most safety proposals focus almost exclusively on risk prevention while neglecting the essential other half of the risk management equation: mitigation. Given AI’s demonstrated utility in medical research, drug development, cybersecurity, materials science, and other critical areas, it is highly likely that future crises—whether AI-related or otherwise—will require rapid AI-powered responses. A federal AI sandbox program would ensure that when such moments arise, the policy infrastructure exists to enable swift, effective, and carefully monitored AI solutions.
Sandbox Risks
Despite these advantages, sandboxes are not risk-free endeavors, and with regulatory flexibility comes opportunities for both abuse and policy failure.
Among the most significant concerns are economic privilege risks. In the wrong hands, these flexible tools could grant favored firms unequal regulatory treatment, which would reduce competition, distort markets, and potentially create government-backed monopolies. Another risk is excessively expansive use. If unbound by appropriate limits, sandboxes could allow an ambitious executive to side-step proper legislative process, enabling unilateral shaping of the regulatory environment.
Another significant challenge lies in the implementation details. Past programs show that sandboxes are high-effort, requiring extensive resources and staff committed to managing applications, monitoring results, ensuring adequate consumer protections, and carefully assessing results to inform regulatory change. Without adequate resources, sandboxes risk acting as policy distractions, wasting the time and resources of both regulators and firms.
Weighing the SANDBOX Act Approach
This short yet far from comprehensive list of potential risks demonstrates that while the upsides can be significant, it is essential that sandbox legislation be designed with thoughtful care. With these potential positives and negatives in mind, let’s return to what the SANDBOX Act framework might mean for the future of AI and AI policy.
In several important respects, this framework is well-positioned to capture innovation benefits. Two important features particularly stand out. First, its breadth and applicability across all regulatory agencies reflect the reality that AI is a general-purpose technology. If we want high-impact AI diffusion across all economic sectors, we’ll need flexibility across all agencies. Second, its technology-agnostic approach to AI is accommodating to exceptionally diverse AI forms, functions, and use cases. Given AI’s rapid evolution, overly prescriptive frameworks risk unfairly favoring certain approaches over others.
Despite these strengths, however, the current framework fails to adequately address the risks outlined above. A fundamental challenge is that the proposed program remains unfunded, imposing expensive new responsibilities on the OSTP and other agencies without providing the resources needed to manage what could be substantial administrative demands.
Scope presents more significant concerns. Regulatory relief duration can extend up to ten years, potentially enabling presidents to grant favored firms a decade of preferential treatment. This offers a pathway to government-enabled market dominance. Moreover, waiver authorities are largely uncapped, offering extraordinary flexibility that could be misused as a fast pass around Congress and appropriate regulatory processes. While flexibility is essential for realizing innovation benefits, problems emerge if sandbox programs become expansive alternatives to formal deregulation. Such misuse would weaken focused study of regulatory impact, undermine program trust, and potentially create privileges so substantial that sandbox participation and its required oversight become economically necessary rather than truly voluntary.
Conclusion
As Congress debates the SANDBOX Act, it is essential that lawmakers consider whether there are ways to address such risks while maintaining the pro-innovation regulatory flexibility that sandboxes can yield. To ensure effective implementation, adequate resources must be discussed and perhaps allocated. To limit potentially excessive use and mitigate economic privilege risks, Congress should also consider placing thoughtful yet still flexible boundaries on both the scope and duration of waivers. As a further safeguard, Congress might also establish a mechanism to allow veto authority over both waiver approvals and denials to prevent potential executive overreach, excess, or favoritism. While a fully comprehensive discussion of all possible modifications to mitigate these risks extends beyond this analysis, further legislative work is clearly needed to ensure program success.
Efforts to provide a deregulatory onramp for innovation through tools like sandboxes can be worthwhile and beneficial. Designed with appropriate safeguards, an AI sandbox could serve as a pragmatic tool to balance rapid AI innovation with necessary oversight and accountability. By enabling this flexibility, Congress can extend US AI leadership and ensure the fast-moving pace of AI innovation extends to all sectors—even the most regulated.