AI Sandboxes Are Shared Ground for US-Mexico Cooperation

Abstract

Artificial intelligence (AI) presents a governance challenge because it evolves rapidly, crosses sectors, and often behaves in context-dependent ways that are difficult to govern through static rules alone. In response, recent U.S. policy proposals have called for regulatory sandboxes to support AI testing and deployment. This report argues that regulatory sandboxes are not simply innovation-friendly exemptions, but institutional mechanisms for supervised experimentation, regulatory learning, and international interoperability. Properly designed sandboxes can help regulators observe AI systems in real-world conditions, identify risks earlier, refine compliance pathways, and improve rulemaking before harms scale.

They can also strengthen U.S. leadership by supporting evidence-based governance and fostering structured cooperation with jurisdictions such as Mexico, where nascent regulatory experimentation initiatives are already beginning to demonstrate the practical value of sandbox approaches and create meaningful opportunities for cross-border collaboration and regulatory alignment.

The report concludes that, if embedded within clear accountability structures, sandboxes can advance both innovation and public protection, while enabling more coordinated and interoperable approaches to AI governance across borders and helping translate broad policy principles into practical regulatory action.

Introduction

The United States has begun to articulate more explicit governance of artificial intelligence, including a call for broader access to testing environments and the use of regulatory sandboxes for AI applications.[1] This recommendation is significant because it recognizes a core reality: AI regulation and leadership requires not just research and investment, but also institutional mechanisms that allow governments to learn from real-world deployment under controlled conditions. Regulatory sandboxes offer one of the most practical ways to meet that need.

This report argues that regulatory sandboxes should be understood as instruments of safe innovation, institutional learning, and global interoperability. Properly designed sandboxes do not merely reduce friction for firms. Rather, they create structured environments in which regulators and developers can jointly generate evidence about how AI systems behave in practice, where risks emerge, and how legal obligations should be calibrated. In that sense, sandboxes help reconcile two goals often treated as competing: accelerating innovation and safeguarding the public interest.[2]

Why AI Governance Requires Sandboxes

AI creates unusual governance problems because it is a general-purpose technology that evolves quickly, operates across sectors, and can produce unpredictable effects when deployed in complex social settings. Traditional regulatory approaches are often too slow or too rigid to respond effectively to these characteristics.[3] Preventive, ex ante rules remain necessary, but they are rarely sufficient when regulators lack timely information about system behavior, data dependencies, downstream uses, and real-world harms.

For that reason, regulatory learning must occur early in the policy cycle. Anticipatory governance emphasizes the need for institutions to build foresight, experimentation, and learning into decision-making rather than waiting for harms to materialize before adjusting rules.[4] Sandboxes operationalize that logic by allowing authorities to observe AI systems in supervised environments before large-scale rollout. In doing so, they help regulators identify blind spots, evaluate alternative compliance strategies, and refine safeguards using evidence rather than assumption.[5]

This function is especially important in AI because some risks are only visible in context. Bias, automation overreach, degraded human oversight, and unintended uses often become clearer through monitored deployment than through laboratory testing or abstract impact assessments alone. Sandboxes therefore offer a way to generate practice-based knowledge that conventional consultation processes cannot fully provide.[6]

Sandboxes as Learning Infrastructures

The value of a regulatory sandbox lies less in temporary flexibility and more in its function as a critical learning infrastructure. A well-designed sandbox can reduce information asymmetries between regulators and firms, improve inter-agency coordination, strengthen internal technical capacity, and create feedback loops between experimentation and rulemaking.[7]

This learning function distinguishes meaningful sandboxes from symbolic pilot programs. If a sandbox only provides firms with a temporary space to test products, but does not produce usable evidence for regulators, it contributes little to governance. By contrast, when sandbox outputs are systematically captured, evaluated, and integrated into administrative or legislative processes, they can improve regulatory quality over time.[8]

This distinction matters for the U.S. While adopting sandboxes simply as deregulatory signals could create confusion or uneven treatment across sectors, building them as institutional learning mechanisms supports more precise, evidence-based governance while preserving accountability.[9] The policy choice is therefore not between regulation and experimentation, but between uninformed regulation and regulation informed by monitored experimentation.

Addressing Concerns About Cost, Safety, and Capture

Critics are right to note that sandboxes are not costless. They require administrative resources, technical expertise, and sustained oversight. Some also argue that uncertainty is not always the main barrier to innovation and that governments should focus instead on reducing unnecessary compliance burdens for already compliant technologies.[10] These concerns should be taken seriously, particularly in resource-constrained agencies.

Even so, the relevant comparison is not between sandboxing and a zero-cost baseline. The more realistic comparison is between the cost of supervised experimentation and the cost of regulating, revising, or litigating after large-scale deployment has already created harm or institutional confusion. When sandboxes are designed as capacity-building mechanisms, they can reduce future regulatory costs by improving rule design, strengthening enforcement feasibility, and limiting the need for reactive, corrective measures.[11]

Concerns about deregulation and regulatory capture are likewise legitimate, but they arise from poor design rather than from the sandbox model itself.[12] Sandboxes become problematic when entry criteria are opaque, monitoring is weak, reporting is minimal, or participants are effectively insulated from accountability. Those risks can be reduced through clear eligibility standards, time limits, transparency requirements, supervisory checkpoints, and termination powers.[13] Properly governed sandboxes are not zones outside the law. They are structured spaces for testing whether existing legal requirements are effective, excessive, insufficient, or misaligned with technological realities.

Interoperability and US Leadership

AI governance is increasingly shaped across borders. Standards, market incentives, transnational value chains, and policy diffusion all influence how national regulatory choices operate in practice. As a result, U.S. leadership in AI governance will depend not only on domestic innovation capacity, but also on its ability to develop governance models that are credible, exportable, and interoperable with other jurisdictions — including in response to the regulatory space that other regions are already claiming.[14]

Regulatory sandboxes can support that goal. Because they are operational rather than purely declaratory, sandboxes allow regulators to compare supervisory models, compliance pathways, and evidentiary practices across jurisdictions. This makes them useful bridging mechanisms between domestic experimentation and international norm development.[15] They can also help avoid a false choice between full harmonization and complete fragmentation. Jurisdictions do not need identical rules to learn from one another; they need compatible methods for experimentation, evaluation, and adaptive adjustment.

For the United States, this has practical strategic value. If U.S. agencies can demonstrate that sandbox-based oversight produces reliable evidence, manages risk credibly, and supports innovation across sectors, they will be better positioned to influence emerging global approaches to AI governance. Leadership, in that sense, depends as much on institutional competence as on policy rhetoric.

Collaboration With Mexican Institutions, Startups, and Researchers

Platforms for Economic Integration

Regulatory sandboxes can also serve as strategic instruments for attracting Mexican AI startups, academic researchers, and applied research teams to collaborate with U.S. agencies. This is especially important in the U.S.-Mexico context, given the high degree of economic integration between the two countries and the growing deployment of AI across shared sectors such as manufacturing, logistics, digital services, finance, and energy. In these areas, sandboxes can provide a practical platform for sustained cooperation between U.S. agencies and Mexican institutions by enabling both countries to exchange evidence on AI system performance, compare supervisory practices, coordinate risk-assessment methods, and identify sector-specific governance needs without requiring full legal harmonization. In this way, sandboxes can strengthen bilateral trust, reduce unnecessary compliance frictions for firms operating across North American value chains, and create a more coherent foundation for cross-border AI governance.[16]

Mechanisms for Institutional Openness

By offering structured testing environments, defined supervisory pathways, and clearer channels for engagement with public authorities, sandboxes also reduce uncertainty for external innovators who might otherwise face high entry barriers when seeking to work with U.S. regulators. For that reason, sandboxes function not only as regulatory tools but also as signals of institutional openness: They show that U.S. agencies are willing to engage emerging developers and researchers in a supervised, evidence-generating process before products scale in the market. This approach is consistent with the White House’s emphasis on expanding access to testing environments needed to build world-class AI systems.[17]

Pathways for Binational Policy Innovation

This could create a particularly valuable pathway for binational cooperation with Mexico. Recent policy discussions on U.S.-Mexico AI relations have explicitly recommended new mechanisms for policy innovation and experimentation, including shared data infrastructures, exchanges of policy best practices, and the development of regulatory sandboxes and policy prototypes involving researchers, policymakers, and civil society from both countries. These arrangements can make collaboration more attractive to Mexican startups and researchers by giving them access to real-world policy dialogue, pilot opportunities, and institutional partners in the United States, while also allowing U.S. agencies to benefit from Mexican expertise, regional experience, and cross-border perspectives on deployment and labor-market effects.[18]

Architecture for Regional Policy Momentum

The opportunity is especially timely because Mexico is already exploring sandbox-based approaches to AI governance. According to the OECD AI Policy Observatory, actors in Mexico have identified the need for an AI sandbox as a flexible mechanism for addressing technical, legal, and ethical challenges, and ongoing efforts have included technical exchange with international partners to strengthen the country’s AI ecosystem.[19] A U.S. sandbox architecture that is open to collaboration with Mexican institutions could therefore channel existing momentum into concrete binational projects — such as joint pilots, researcher exchanges, comparative evaluations, and shared governance protocols — without requiring full legal harmonization. Properly designed, such arrangements could help the U.S. attract high-potential Mexican innovators while simultaneously strengthening regulatory learning and bilateral trust.[20]

Policy Design Implications

As the U.S. expands its use of AI regulatory sandboxes, several design principles should guide their implementation.

1. Clear objectives: Sandbox programs should be tied to clear public objectives, not simply to generalized support for innovation. Agencies should specify what they are trying to learn, what risks they are monitoring, and how findings will inform future decisions.[21]
2. Integrated governance: Sandboxes should be embedded in regulatory processes rather than treated as stand-alone experiments. This means linking participation criteria, evaluation metrics, and reporting structures to actual administrative learning and future rulemaking.[22]
3. Credible oversight: Oversight mechanisms should be visible and credible. Transparency, supervision, and exit powers are essential for public trust and legitimacy.[23]
4. Strategic interoperability: Sandbox design should be attentive to interoperability. AI systems often operate across sectors and borders, so the U.S. should favor models that facilitate cooperation with peer regulators, standards bodies, and trusted foreign partners where appropriate.[24] A sandbox that generates evidence only for a single isolated agency may still be useful, but a sandbox that also supports inter-agency and cross-border learning will be far more valuable.

Conclusion

The case for AI regulatory sandboxes is ultimately a case for more informed governance. The White House’s 2026 legislative recommendations correctly recognize that access to testing environments is necessary for American leadership in AI.[25] But sandboxes should not be framed simply as tools for reducing barriers to innovation. Their real significance lies in their capacity to produce evidence, build regulatory capability, and support more adaptive forms of public oversight.

When designed well, sandboxes help move governance from static rulemaking toward monitored experimentation, iterative learning, and institutional adaptation. They can improve safety without freezing innovation, and they can strengthen innovation without abandoning accountability. For the United States, regulatory sandboxes are not a peripheral policy option, but a central instrument for the next phase of AI governance.

Notes

[1] The White House, A National Policy Framework for Artificial Intelligence, March 2026, https://www.whitehouse.gov/wp-content/uploads/2026/03/03.20.26-National-Policy-Framework-for-Artificial-Intelligence-Legislative-Recommendations.pdf.

[2] Hilary J. Allen, “Regulatory Sandboxes,” George Washington Law Review 87 (2019), available at SSRN, https://doi.org/10.2139/ssrn.3056993; Angela Attrey et al., “The Role of Sandboxes in Promoting Flexibility and Innovation in the Digital Age,” OECD Going Digital Toolkit Notes, No. 2, OECD Publishing, https://www.oecd.org/content/dam/oecd/en/publications/reports/2020/06/the-role-of-sandboxes-in-promoting-flexibility-and-innovation-in-the-digital-age_ddcd3d40/cdf5ed45-en.pdf; and Thomas Buocz et al., “Regulatory Sandboxes in the AI Act: Reconciling Innovation and Safety?” Law, Innovation and Technology 15, no. 2 (2023): 357–89, https://doi.org/10.1080/17579961.2023.2245678.

[3] David H. Guston, “Understanding ‘Anticipatory Governance’,” Social Studies of Science 44, no. 2 (2014): 218–42, https://doi.org/10.1177/0306312713508669; Matthijs M. Maas, “Aligning AI Regulation to Sociotechnical Change,” in The Oxford Handbook of AI Governance edited by Justin B. Bullock et al. (Oxford University Press, 2022), 358–80, https://doi.org/10.1093/oxfordhb/9780197579329.013.22; and Stefan Seidel et al., “Regulating Emerging Technologies: Prospective Sensemaking Through Abstraction and Elaboration,” MIS Quarterly 49, no. 1 (2025): 179–204, https://misq.umn.edu/misq/article/49/1/179/88/Regulating-Emerging-Technologies-Prospective.

[4] Deirdre Ahern, “The New Anticipatory Governance Culture for Innovation: Regulatory Foresight, Regulatory Experimentation and Regulatory Learning,” European Business Organization Law Review, 26, no. 2 (2025): 241–83, https://doi.org/10.1007/s40804-025-00348-7; Harry Armstrong et al., “Renewing Regulation: Anticipatory Regulation in an Age of Disruption,” Nesta, March 2019, https://media.nesta.org.uk/documents/Renewing_regulation_v3.pdf; and Piret Tõnurist and Jack Orlik, “Towards Anticipatory Governance Guidelines for Public Sector Organisations,” OECD Working Papers on Public Governance no. 82, OECD Publishing, May 22, 2025, https://doi.org/10.1787/a5203d0b-en.

[5] Lori Bennear and Jonathan B Wiener, “Adaptive Regulation: Instrument Choice for Policy Learning over Time,” Mossavar-Rahmani Center for Business and Government, Harvard Kennedy School, working paper, February 12, 2019, https://bit.ly/3QzhRke; Urs Gasser and Viktor Mayer-Schönberger, “On the Shoulders of Others: The Importance of Regulatory Learning in the Age of AI,” Virginia Journal of Law & Technology, 28, no. 1 (2024): 1–14, https://bit.ly/4d2s1BI.

[6] Gabriel Kwok Hui Chen and Araz Taeihagh, “Designing Regulatory Sandboxes: A Comprehensive Framework for Aligning Functionalities and Objectives,” Policy Design and Practice 9, no. 1 (2026): 1–15, https://doi.org/10.1080/25741292.2025.2570954; “AI Regulatory Sandbox Mexico,” OECD.AI, updated December 24, 2025, https://oecd.ai/en/dashboards/policy-initiatives/ai-regulatory-sandbox-mexico-4522.

[7] Armando Guio Español, “Regulatory Sandboxes in Developing Economies: An Innovative Governance Approach,” Economic Commission for Latin America and the Caribbean, July 19, 2024, https://www.cepal.org/en/publications/80496-regulatory-sandboxes-developing-economies-innovative-governance-approach; Kaia Kert et al., “Regulatory Learning in Experimentation Spaces,” European Commission, December 12, 2022, https://publications.jrc.ec.europa.eu/repository/handle/JRC130458.

[8] “The New Anticipatory Governance Culture for Innovation: Regulatory Foresight, Regulatory Experimentation and Regulatory Learning”; Organisation for Economic Co-operation and Development (OECD), “Regulatory Experimentation: Moving Ahead on the Agile Regulatory Governance Agenda,” OECD Publishing, April 5, 2024, https://doi.org/10.1787/f193910c-en; and David Vogel, “Regulatory Excellence: The Role of Policy Learning and Reputation, Penn Program on Regulation,” University of Pennsylvania, June 2015, https://pennreg.org/regulatory-excellence/wp-content/uploads/sites/5/2023/01/vogel-ppr-bicregulatordiscussionpaper-06-2015.pdf.

[9] Paulo Carvão, “Are AI Regulatory Sandboxes a Good Idea?” Forbes, September 10, 2025, https://www.forbes.com/sites/paulocarvao/2025/09/10/are-ai-regulatory-sandboxes-a-good-idea/.

[10] Dan Quan, “A Few Thoughts on Regulatory Sandboxes,” Stanford Center on Philanthropy and Civil Society, September 25, 2019, https://pacscenter.stanford.edu/a-few-thoughts-on-regulatory-sandboxes.

[11] Financial Conduct Authority, Regulatory Sandbox Lessons Learned Report, October 2017, https://www.fca.org.uk/publication/research-and-data/regulatory-sandbox-lessons-learned-report.pdf; Ivo Jeník and Schan Duff, How to Build a Regulatory Sandbox: A Practical Guide for Policy Makers, World Bank Group, September 1, 2020, https://documents.worldbank.org/en/publication/documents-reports/documentdetail/126281625136122935; and OECD, “Regulatory Sandbox Toolkit,” OECD Publishing, July 24, 2025, https://doi.org/10.1787/de36fa62-en.

[12] “Are AI Regulatory Sandboxes a Good Idea?”

[13] Claudio Novelli et al., “Getting Regulatory Sandboxes Right: Design and Governance Under the AI Act,” SSRN, June 30, 2025, https://doi.org/10.2139/ssrn.5332161; Sofia Ranchordás, “Experimental Regulations and Regulatory Sandboxes: Law Without Order?” Law & Method, https://doi.org/10.5553/REM/.000064; and Ranchordás and Valeria Vinci, “Regulatory Sandboxes and Innovation-Friendly Regulation: Between Collaboration and Capture,” Italian Journal of Public Law 1 (2024), https://doi.org/10.2139/ssrn.4696442.

[14] Ben Crum, “Brussels Effect or Experimentalism? The EU AI Act and Global Standard-Setting,” Internet Policy Review 14, no. 3 (2025): 2–21, https://doi.org/10.14763/2025.3.2032; Emmanouil Papagiannidis et al., “Responsible Artificial Intelligence Governance: A Review and Research Framework,” The Journal of Strategic Information Systems 34, no. 2 (2025): 101885, https://doi.org/10.1016/j.jsis.2024.101885; and Charlotte Siegmann and Markus Anderljung, “The Brussels Effect and Artificial Intelligence: How EU Regulation Will Impact the Global AI Market,” Centre for the Governance of AI, August 2022, https://cdn.governance.ai/Brussels_Effect_GovAI.pdf.

[15] “Regulatory Sandboxes in the AI Act: Reconciling Innovation and Safety?”; “Designing Regulatory Sandboxes: A Comprehensive Framework for Aligning Functionalities and Objectives.”

[16] Guio Español et al., “The Importance and Challenges of Developing a Regulatory Agenda for AI in Latin America,” in Elgar Companion to Regulating AI and Big Data in Emerging Economies, edited by Mark Findlay et al. (Edward Elgar Publishing, 2023), 201–27, https://doi.org/10.4337/9781785362408.00019; Eduardo Levy Yeyati, “An Enabling Regulatory Framework for Artificial Intelligence in Latin America and the Caribbean,” Inter-American Development Bank, November 2025, http://dx.doi.org/10.18235/0013801; and Andrés Mosqueira and Shaanty Emmanuel Rubio Gonzalez, “Foster Innovation or Mitigate Risk? AI Regulation in Latin America,” White & Case, November 18, 2024, https://www.whitecase.com/insight-our-thinking/latin-america-focus-2024-ai-regulation.

[17] A National Policy Framework for Artificial Intelligence.

[18] Rodrigo Ferreira et al., “AI and US-Mexico Relations: The Future(s) of Work,” Rice University’s Baker Institute for Public Policy, May 28, 2025, https://doi.org/10.25613/SST6-3T58.

[19] AI Regulatory Sandbox Mexico,” OECD.AI, updated December 24, 2025, https://oecd.ai/en/dashboards/policy-initiatives/ai-regulatory-sandbox-mexico-4522?; Ernesto Ibarra et al., “Panorama de la Inteligencia Artificial en México: Hacia una Estrategia Nacional,” March 2024, https://www.amcid.org/page/sandboxregulatoriomexico.

[20] “AI Regulatory Sandbox Mexico.”

[21] “Designing Regulatory Sandboxes: A Comprehensive Framework for Aligning Functionalities and Objectives”; “Regulatory Sandbox Toolkit.”

[22] “The New Anticipatory Governance Culture for Innovation: Regulatory Foresight, Regulatory Experimentation and Regulatory Learning”; “Regulatory Learning in Experimentation Spaces.”

[23] Financial Conduct Authority, Regulatory Sandbox Lessons Learned Report, October 2017, https://www.fca.org.uk/publication/research-and-data/regulatory-sandbox-lessons-learned-report.pdf; “Getting Regulatory Sandboxes Right: Design and Governance Under the AI Act.”

[24] “Brussels Effect or Experimentalism? The EU AI Act and Global Standard-Setting.”

[25] A National Policy Framework for Artificial Intelligence.

This publication was produced by Rice University’s Baker Institute for Public Policy. Wherever feasible, the material was reviewed by outside experts prior to release. Any errors or omissions are solely the responsibility of the author(s).

This material may be quoted or reproduced without prior permission, provided appropriate credit is given to the author(s) and Rice University’s Baker Institute for Public Policy. The views expressed herein are those of the individual author(s) and do not necessarily represent the views of Rice University’s Baker Institute for Public Policy.