2.1. Regulatory sandboxes as innovation incubators

We follow Johnson (Reference Johnson2022, p.4) in understanding regulatory sandboxes as ‘a distinct type of regulatory instrument within the larger categories of flexible, experimentalist, and anticipatory modes of governance’. They are characterized by criteria for entry, restricted discretion by the regulator, rules for risk management, an orientation towards certain substantive success criteria, and a structured exchange between regulators and the objects of regulation (Johnson, Reference Johnson2022, pp. 5–7). Regulatory sandboxes are an increasingly widespread form of regulatory experimentation that governments employ for a temporary live testing of technological innovations under increased oversight.

This form of experimental regulation has been gaining lots of momentum within less than a decade. After the UK’s Financial Conduct Authority established the first regulatory sandbox for financial technologies (Fintech) in 2015, many regulators over the globe have embraced this instrument to test technological innovation and gain evidence that can inform policymaking. Whereas a World Bank report from 2020 covered experiences from 73 fintech sandboxes in 57 countries (2020), a study by Markellos et al. (Reference Markellos, Ennis, Enstone, Manos, Pazaitis and Psychoyios2024) mapped 199 cases across 92 countries, covering various sectors such as telecommunications, energy, and AI. These numbers are likely to increase further especially with an increasing interest in regulatory sandboxes for AI.

Major possible beneficial impacts of these sandboxes are promoting innovation, greater legal certainty for businesses, and increased knowledge for policymakers, possibly leading to adapted regulation (Ranchordás, Reference Ranchordás2021b). Policymakers place strong hopes in regulatory sandboxes in view of their potential to ‘improve the regulatory governance of innovation and accelerate the deployment of innovative solutions’ (Ranchordás & Vinci, Reference Ranchordás and Vinci2024, p. 2). It is notable that references to the objectives of regulatory sandboxes commonly mention innovation and access to markets first (e.g., Buocz et al., Reference Buocz, Pfotenhauer and Eisenberger2023; OECD, 2023; Ranchordás, Reference Ranchordás2021b), whereas learning and producing insights as an important aspect of sandboxes seem to be subordinate (see also Ruschemeier, Reference Ruschemeier and Steffen2025). This is also palpable in statements that frame learning within a regulatory sandbox as instrumental for fostering innovation. For instance, as Nabil (Reference Nabil2024, p. 244) notes: ‘At their best, regulatory sandboxes can promote technological innovation by attracting innovative companies and helping policymakers design an evidence-based, iterative approach to regulating emerging technologies’. In this quote, producing evidence is not so much a goal in itself, but rather a means to supporting innovation.

A focus on the objective of promoting innovation and business creation is also recognizable in empirical research interested in the effects of regulatory sandboxes. The jury is still out, however, on whether sandboxes actually realize this objective, with available evidence still being scarce and showing a mixed picture. An evaluation of the fintech sandbox created in the UK finds that the sandbox increased the capital raised by participants by 15% while the probability to increase capital increased by 50% (Cornelli et al., 2024). Another study on the same regulatory sandbox, however, concludes that this regulatory instrument has negatively impacted the financial performance of digital banks due to higher compliance and efficiency costs (Washington, Rehman & Lee, Reference Washington, Rehman and Lee2022). Hellmann et al. (Reference Hellmann, Montag and Vulkan2024) also revisited the previous analysis of the UK fintech sandbox and showed that when using statistical estimation techniques that account for sandboxes having heterogeneous effects on acquired funding, the overall positive relationship is attributable to a single sandbox. In their model, specifying heterogeneous treatment effects, the overall contribution of regulatory sandboxes to raising capital becomes statistically insignificant. At the same time, the authors do point to indirect effects of a sandbox on the broader fintech ecosystem, as the entry into a sandbox seems to increase the funds raised by other high-growth companies operating in the same industry sector (Hellmann et al., Reference Hellmann, Montag and Vulkan2024).

Mixed evidence also results from further studies that have looked at more than one country. A study comparing the fintech sandboxes in the UK and in Singapore, in turn, yields no clear evidence that they raised more capital in the first 2 years (Chen, Reference Chen, Fenwick, Vyn Uytsel and Ying2020). A cross-country analysis with overall 18 countries finds that regulatory sandboxes have led to an increased influx of venture capital into the fintech venture ecosystem (Goo & Heo, Reference Goo and Heo2020). The authors arrived at this finding by matching nine countries that have introduced a regulatory sandbox with nine countries that had no sandbox based on similar National Competitiveness Index scores. Evidence on regulatory sandboxes is furthermore limited because it comes from the realm of fintech. An important exception is the study by Beckstedde et al. (Reference Beckstedde, Correa Ramírez, Cossent, Vanschoenwinkel and Meeus2023), which analyzes the experiences from regulatory sandboxes in the energy sectors of eight European countries. The authors conclude that regulatory derogations used in the studied sandboxes ‘validate the idea of using regulatory sandboxes to promote innovation’. Furthermore, the authors argue based on their findings that the regulatory scope of sandboxes should be as open as possible in order to best realize the goal of promoting innovation.

Overall, it appears that promoting innovation and new business models is the ultimate standard by which the success of sandboxes is judged. If they fail to foster innovation, they do not appear to merit the effort. However, promoting innovation is a difficult standard since innovation is ‘an elusive concept which is hard to define, measure, and thus regulate’ (Ranchordás & Vinci, Reference Ranchordás and Vinci2024, p. 15). Given that it is uncertain how to reliably support innovation, it seems unclear how a regulatory sandbox can contribute to it in a targeted, predictable fashion. At the same time, designing the sandbox to foster innovation may invite concerns about regulators being involved in the innovation process themselves. A notable example of this is how certain authorities in the United States have interpreted this methodology. In 2018, Securities and Exchange Commissioner Hester Peirce expressed concerns regarding the design and implementation of regulatory sandboxes, particularly the idea of regulatory authorities actively participating in the innovation process. According to Peirce, sandboxes risk becoming spaces where regulators exert influence over the design of a technology, thereby limiting innovation and restricting the ability to propose new models (Pierce, Reference Peirce2018). A related critical assessment of regulatory sandboxes comes from Germany’s Federal Financial Supervisory Authority (BaFin), which has argued against the use of sandboxes because their goal of promoting innovation would not fall under their mandate and would not require the authority’s involvement (Ruschemeier, Reference Ruschemeier and Steffen2025, p. 328).

However, these perspectives present sandboxes as a way to create a space in which the design of a technology and the innovation process itself is debated. For instance, in the case of AI, this interpretation assumes that regulatory authorities would play a direct role in shaping the operation or training of these models. This is hardly a suitable mandate for regulators. Indeed, regulatory authorities do not possess the needed expertise and technical capabilities to engage in such interactions. In fact, regulators are struggling to keep pace with the rapid evolution of technology. Rather than seeking to shape innovation directly, many authorities view regulatory sandboxes as mechanisms for learning – allowing them to better understand emerging technologies and their implications for regulatory frameworks.

2.2. Regulatory sandboxes as a tool to accelerate learning within the policy cycle

The design and governance of a sandbox cannot guarantee that it leads to more innovation. A regulatory sandbox can, however, be systematically created and operated in a way that it contributes to learning processes which can help to adapt to changing conditions, specifically through technological developments. Regulatory sandboxes are a form of adaptive or agile regulation that is attuned to conditions of ongoing and rapid technological change. Such agile regulation is proactive through anticipating possible consequences, it is responsive through updating rules quickly and dynamically, and it is based on the systematic use of evidence, which enables adaptivity and learning (Allen, Reference Allen2019; Guston, Reference Guston2014; Howlett et al., Reference Howlett, Capano and Ramesh2018; Ranchordás, Reference Ranchordás2021a, Reference Ranchordás2021b). In this way sandboxes can help with evaluating the suitability of legal frameworks, save time through advancing learning processes, and inform policymaking (BMWi, 2019). Emphasizing this adaptive dimension of regulatory sandboxes means shifting the focus from promoting innovation to fostering learning processes through sandboxes.

Empirical evidence on the impacts of regulatory sandboxes in terms of the learning they engender is still very scarce, but available case studies indicate that these regulatory tools increase knowledge and capacities. Based on their interviews from regulators involved in sandboxes from five jurisdictions, Alaassar et al. (Reference Alaassar, Mention and Aas2020) conclude that the interaction between regulators and regulates improves their understanding of the constraints of existing regulation, technological risks, and regulatees’ need for support. Macrae and Ansell (Reference Macrae and Ansell2024) further conclude from their ethnographic case study of a regulatory sandbox for AI applications in medical diagnostic services implemented in the UK that the sandbox establishes a generative space for collaborating, learning, and innovation. The authors stress that the regulatory sandbox created an adaptive space for exploring and learning and that the commitment to learning was rooted in an awareness of organizational ignorance and a need to close regulatory knowledge gaps (Macrae & Ansell Reference Macrae and Ansell2024, pp. 20–21).

The flexible, dynamic, and open approach of a sandbox is diametrically opposed to the idea that policymakers and regulators are generally in a position to set rules which are suitable for governing technological innovations. Rather, regulatory sandboxes as a form of agile regulation are an expression of ‘regulatory humility’ as described by Dunlop and Radaelli (Reference Dunlop and Radaelli2016), i.e., the recognition of the contingency of outcomes following from regulatory choices and a reflection on the limits of policymakers’ and regulators’ control. As Haeffele and Hobson (Reference Haeffele and Hobson2019, p. xiii) note, policies can fall short of realizing the intended effects due to an unacknowledged contingency and complexity of problems, making humility in regulation a healthy and prudent approach. This also means that regulators and policymakers must themselves embrace the role of learners to make the experimenting mode of regulatory sandboxes work – which can be unfamiliar and a source of discomfort among regulators (Macrae & Ansell Reference Macrae and Ansell2024, p. 21).

Unlike the objective of promoting innovation, emphasizing the importance of learning is vital in navigating complexity, uncertainty, and volatility, as such conditions necessitate quicker learning and broader exploration of options. Technological innovations with major transformative impacts on society can significantly contribute to conditions that are volatile and marked by increased uncertainty. These conditions present special challenges for regulators, as they will lack knowledge regarding how innovations will affect markets and society. Technological innovations may also span different areas and blur lines between markets and sectors (e.g., the automotive industry and mobility services), and the high pace of change may quickly make extant regulatory definitions and categories become obsolete. The governance of emerging technologies thus calls for more flexibly responding to regulatory problems that these technologies may pose (Stilgoe et al., Reference Stilgoe, Owen and Macnaghten2013). To develop sound policies, as Sunstein (Reference Sunstein2022) notes, regulators must go beyond relying solely on expert judgment and instead prioritize systematic learning and real-time evidence gathering. The objective should be to assess the real-world impact of policies in advance, rather than retrospectively, allowing for more adaptive and informed decision-making (Sunstein, Reference Sunstein2022).

These considerations apply particularly to the regulation of AI systems. AI is a general-purpose technology that can lead to new opportunities and risks in manifold areas and across different sectors (Ruschemeier, Reference Ruschemeier2023). The complexity and opacity of AI applications leads to information asymmetries, with regulators struggling to understand the implications of AI uses. Innovations in AI have also been occurring at a fast pace in recent years, leading to uncertainty about the suitability of regulation and possibly legal gaps. They thus create an acute need for learning and gaining a better understanding whether and how certain rules may or may not be suitable – as they can hinder certain socially beneficial innovations or may leave important novel sources of risk unaddressed.

Two important characteristics of regulatory sandboxes are particularly important for facilitating the kind of learning that can help to close legal gaps regarding AI regulation. First, through establishing collaboration between stakeholders, they can contribute to reducing information asymmetries. Close interactions between private sector actors and regulators can play an important role in meeting information needs of regulators, reducing knowledge asymmetries and fostering the necessary understanding of technologies (Fahy, Reference Fahy2022; Parenti, Reference Parenti2020). As Sunstein (Reference Sunstein2022) highlights, when regulators face significant gaps in information – often referred to as ‘the knowledge problem’ – they can benefit from tools that harness expertise from the private sector. A well-designed regulatory system should aim to bridge this gap, including through mechanisms of advanced testing to generate insights before formal regulatory decisions are made.

Some scholars have been vocal about the potential use of sandboxes as instruments of regulatory capture (e.g., Ranchordás & Vinci, Reference Ranchordás and Vinci2024). However, the status quo is already one of a concentration of technical knowledge that puts regulators at a disadvantage and may create an environment that AI developers actively seek. Specifically, under circumstances that can be described as too complex to regulate (Neuman, Reference Neuman2023), the result may well be limited regulatory oversight due to the inherent complexity of the technology and the knowledge asymmetries between innovators and regulators. Furthermore, the concrete governance of a regulatory sandbox is crucial for preventing undue industry influence. Openness and actively engaging various stakeholders, for instance, can reduce the risk of regulatory capture (Parenti, Reference Parenti2020). Ultimately, to achieve their intended impact, it is essential to design and implement regulatory sandboxes effectively, ensuring they facilitate meaningful learning and robust collaboration. Their design determines the extent to which they can mitigate risks, including consumer harms, regulatory capture, and market distortions, ultimately shaping the effectiveness of regulatory adaptation.

A second characteristic of regulatory sandboxes that is crucially important for promoting learning is that they can involve proactive testing and trialling not only of technology, but also of rules for regulating the technology. Again, the extent to which a sandbox is geared towards such experimentation is a question of its concrete design and governance (Bijkerk, Reference Bijkerk2021). Through directly supervising the trialling of technology in a controlled setting, regulators can gain hands-on experience and knowledge necessary to evaluate how certain rules may be suitable or not to govern AI applications. Within a sandbox, it becomes possible to test technology and also to experiment with and adapt rules, gaining insights about their suitability. In this regard, regulatory sandboxes have strong parallels to the long-standing practice of clinical trials (Buocz et al., Reference Buocz, Pfotenhauer and Eisenberger2023) and to real-world laboratories that have been used in research stressing experimentation, evaluation, and learning (Bergmann et al., Reference Bergmann, Schäpke, Marg, Stelzer, Lang, Bossert, Gantert, Häußler, Marquardt, Piontek, Potthast, Rhodius, Rudolph, Ruddat, Seebacher and Sußmann2021).

The characteristics described above make regulatory sandboxes different from traditional regulation following the policy cycle. The long-term horizon of learning in the policy cycle and the idea of at least somewhat rational and linear policy change are at odds with the open learning process of regulatory sandboxes in which regulatory frameworks are tested and experimented with. Traditional regulation also tends to lag behind technological developments as it is based on the given state of knowledge, and updating regulation will only occur after waiting and seeing what the impacts of technologies are. As Ringe and Ruof (Reference Ringe and Ruof2020, p. 617) note ‘[i]n regular circumstances, the regulator or legislature needs to assess risks on the market. By the time the findings of this process reach the legislative stage, the stakes are usually high, or risks are already on the verge of materializing’.

Regulation along the lines of the model of the common policy cycle thus easily fails to accommodate the pace of technological change in certain instances. While policy learning does occur via the policy cycle, it may simply be too slow. Regulatory sandboxes establish a learning process within a very different setting and timeframe. Yet, although they do not conform to the policy cycle, they can, however, very much complement the traditional policy cycle by accelerating learning processes within it. Regulatory sandboxes essentially advance the process of arriving at new relevant insights, rather than waiting for impacts to show and taking them as an occasion to update existing or create new rules. Regulatory sandboxes can thus operate as policy learning accelerators that form an integral part of the policy cycle.

When gaining insights about the suitability of regulation, an important factor concerning AI applications is that there will commonly not exist any binding rules specifically for regulating AI. To date, only a few jurisdictions, like the EU or China, have adopted binding legislation concerning the development, deployment, and use of AI systems. Depending on which circumstances hold true, regulatory sandboxes for AI can complement the policy cycle in two different ways.

First, regulatory sandboxes can be used to complement existing binding rules concerning AI. In this case, they are an instrument that introduces flexibility and adaptivity to compensate for the rigidity of extant binding regulation. This scenario corresponds to the European Union’s approach to adopting regulatory sandboxes for AI. Article 57 of the EU AI Act specifies what kind of learning is envisaged with these sandboxes. The learning object is evident in one of the intended main outcomes of these sandboxes: the production of results reports that present the key learning outcomes from each initiative. Notably, paragraph 9 of the same Article 57 states that learning is not only intended for entrepreneurs or participating tech companies but also for regulatory authorities. This is emphasized by the statement that one of the objectives of sandboxes created under this article is ‘evidence-based regulatory learning’ (own emphasis).

The European Directorate-General for Research and Innovation defines regulatory learning as the collection and use of any evidence or knowledge relevant to current or future regulatory policy, generated in the process of experimenting with an innovative solution (Molinari, Rachinger, Kordel, Hennen & Van Roy, Reference Molinari, Rachinger, Kordel, Hennen and Van Roy2022). The collection of this type of evidence has been made explicit in the process of designing sandboxes in Europe, such as in the cases of Spain and Germany. The AI sandbox project led by the Secretariat of State for Digitalization and Artificial Intelligence in Spain has emphasized the importance of practical learning so that authorities can support the development of standards, guidance, and tools at both the national and European levels (SEDIA, 2025). As a regulatory instrument that is aligned with the Spanish digital strategy, España Digital 2025, the AI sandbox is a standardized framework and controlled testing environment ‘designed to support innovation while ensuring compliance with the AI Act, particularly focusing on high-risk AI systems’ (Lumsa, 2025).

In Germany, efforts have focused on the development of experimentation clauses that facilitate learning through sandbox environments. The objective is to balance the ability to experiment with compliance with existing legal frameworks, thus achieving regulatory flexibility that makes sandboxes functional and efficient (Otter, Reference Otter2024). The German government has made the development of experimentation clauses a priority within its sandbox policy, emphasizing their role in enabling legislators to gain early insights into innovations, assess their real-world impacts, and identify appropriate legal frameworks. These insights, in turn, support the evolution of broader regulatory systems (BMWK, 2024). Notably, German authorities stress that experimentation clauses serve a dual purpose: (i) facilitating regulatory experimentation and (ii) promoting regulatory learning (BMWK, 2024, p. 11).

Under the EU regulatory framework, it is thus possible to achieve flexibility that sustains regulatory learning. However, this flexibility and learning occur against the backdrop of a comprehensive EU regulation of AI, which most comprehensively regulates AI systems in the high-risk class for which reducing uncertainty is especially important. Furthermore, the possibilities of experimenting within a regulatory sandbox are limited by the fact that the AI Act does not specify possibilities of exemptions from applicable legislation (Buocz et al., Reference Buocz, Pfotenhauer and Eisenberger2023) and that sandbox participants remain fully subject to liability (Truby et al., Reference Truby, Brown, Ibrahim and Parellada2022). In the EU regime of regulatory sandboxes for AI, these are thus mainly a way of testing how existing rules fare with new technological developments and innovations particularly in the domain of high-risk applications.

A different, second constellation exists where regulatory sandboxes are established in the absence of any binding regulation specifically for AI. The sandboxes then serve the purpose of producing evidence that can be used to address the lack of a legal framework for AI governance. They speed up and advance the process of understanding the impacts of the technology and possible rules for governing before governments and regulators take action to adopt binding rules. The role of learning is thus a different one compared to the first scenario described above. It is less constrained by existing regulation and can embrace greater openness and flexibility in the process of obtaining insights that can inform regulatory action. The need for learning is also arguably more acute given that regulation is lacking. One thus has to look beyond the EU to find examples of this second case. In the next section, we will thus examine several cases of regulatory sandboxes from different continents and assess them with regard to their role in engineering policy learning.

3.1. Case selection

To shed light on how regulatory sandboxes for AI are designed to promote policy learning under conditions when AI regulation is lacking, we draw on case studies from the so-called majority world, i.e., outside the EU, North America, and Australia. Countries in that part of the world have lower technological capabilities and regulatory capacities. They also lack the market power of the EU that can compel foreign businesses to play by domestic rules. These conditions make them more sensitive to adopting regulation that could be overly comprehensive and strict, and thus stifle innovation and deter investments. Selecting relevant cases is not a straightforward task as regulatory sandboxes can be a form of window-dressing to signal innovativeness. Therefore, we have chosen to focus on examples from diverse countries that meet the following three criteria. First, the cases reflect established commitments or implemented regulatory sandbox as documented in official reports, public policies, or international organization publications. Second, they have been explicitly proposed as tools to foster the development of regulatory sandboxes. Third, they offer insights into the key motivations behind their creation and the objectives pursued by the authorities and organizations promoting them. Based on these considerations, we have selected the cases of regulatory sandboxes for AI shown in Table 1.

Table 1.

Selected cases of regulatory sandboxes for AI

The regulatory sandboxes planned or implemented in Brazil, Colombia, Mauritius, Mexico, Rwanda, and Thailand fulfil the criteria outlined above. They are a selection of diverse cases from countries with different legal systems and innovation capabilities. In the analysis below, these cases illustrate to what extent learning has become a priority for authorities worldwide in the development of regulatory sandboxes. The analysis is based on a reading of the publicly accessible documentation of the selected regulatory sandboxes (for an overview of the sources used in the analysis, see Appendix A1). This material was analyzed with regard to the conceptualization and practical application of learning, specifically how the various sandboxes institutionalize the testing of technologies and regulatory rules and establish collaboration between stakeholders. Based on a careful reading of the available material, we identify larger themes and point to commonalities and differences of the examined cases. In this way, the analysis offers insights into the regulatory adaptability achieved through sandbox frameworks across different legal, institutional, and cultural contexts.

3.2. Analysis

Mutual learning as a core objective. One of the fundamental expectations tied to the examined regulatory sandboxes is mutual learning. While authorities aim to gain insights from these experimental frameworks, they do not position them solely as learning opportunities for regulators. Instead, selected participants are also expected to benefit from the process. In certain jurisdictions, such as Colombia and Mauritius, authorities explicitly recognize the need to deepen their understanding of emerging technologies by engaging directly with industry actors and observing their operational dynamics. In Thailand, the establishment of testing parameters is envisioned as a collaborative process, albeit guided by an initial proposal from the regulatory authority. In this way, rather than merely fostering innovation, these cases suggest that the authorities seek to understand the types of technologies being developed in their countries and generate insights that complement their AI regulatory initiatives.

Among the jurisdictions emphasizing learning, Mauritius stands out for its explicit adoption of a learning by doing approach. This strategy enables regulatory authorities to acquire a deeper understanding of advanced AI systems and their implications for supervisory responsibilities. By actively engaging with AI-driven technologies in real-world settings, regulators in Mauritius aim to enhance their capacity to oversee and shape the development of these systems effectively. Cases such as those of Colombia and Brazil furthermore clearly illustrate that learning is not a one-way process. The participants developing these innovations can also learn from regulators. In these instances, it becomes evident that some participants may lack awareness regarding the application of regulations in the development or deployment of these technologies. Regulatory authorities, therefore, seek to impart this knowledge and share insights to ensure compliance.

The case of Colombia in particular emphasizes the contribution of mutual learning to facilitate the responsible use of AI. The Colombian AI sandbox was initiated to demonstrate how privacy-by-design principles could enable novel AI solutions while adhering to the country’s data protection regulations. In this setting, the process of mutual learning is again a shared interest. Supervisory agencies can gain a deeper understanding of innovation, while technology developers can engage more effectively with regulators and benefit from their capacity for regulatory innovation. Two ways of innovation are being shared and learned.

This emphasis on mutual learning is very much in line with conclusions that Ringe and Ruof (Reference Ringe and Ruof2018) derived from their analysis of the EU regulatory framework governing so-called robo-advisors, automated financial advisory services operating without human intervention. The authors recommend the use of a ‘guided sandbox’ to foster mutual learning between firms and regulators, thereby also reducing regulatory uncertainty for participants. Overall, learning processes and particularly mutual learning are a central part in the design of the examined regulatory sandboxes. Furthermore, these sandboxes facilitate the learning process based on rules that allow at least for a certain degree of flexibility and experimenting within the setting of a regulatory sandbox.

Legal exemptions and flexibility provided to promote learning. Among the selected cases, it is particularly interesting to observe how regulatory authorities seek to establish mechanisms for regulatory flexibility and to gather valuable insights from these spaces. Across the different cases identified, there is a clear intent to provide legal exemptions that facilitate the ability of both entities and innovators to deploy technology and experiment with new compliance mechanisms. In this regard, three key types of mechanisms emerge: (i) flexible legal interpretations; (ii) regulatory flexibility granted through legal authorization; and (iii) legal flexibility conferred by authorities with the discretion to do so – with the latter corresponding the most to the EU’s approach to regulatory sandboxes, which does not specify exemption mechanisms and gives considerable legal flexibility to member state authorities in designing their sandboxes.

Notably, there is no single model for implementing mechanisms that introduce flexibility. Authorities often seek input from participants to determine the most suitable approach, as seen in cases such as Thailand and Mauritius. In contrast, countries like Colombia and Brazil have leveraged their ability to interpret existing legal frameworks to test different regulatory possibilities, thereby facilitating the process without requiring a pre-existing rule that explicitly authorizes such experimentation. In Mexico and Rwanda, in turn, the expectation is that the government will create dedicated spaces and define the norms or guidelines under which experimentation can take place. Undoubtedly, some authorities feel they have the necessary capabilities to manage the liabilities they may face from administering and overseeing a regulatory sandbox, while others believe they require greater legal support or prefer to delegate this responsibility to the participants.

The role of reporting mechanisms in knowledge transmission and standards setting. A crucial element in facilitating knowledge transfer within regulatory sandboxes is the systematic reporting of results and insights gathered throughout the process. In Thailand, the AI sandbox initiative places significant emphasis on structured reporting mechanisms, specifying the type of information that regulatory authorities expect from participants. Similarly, Brazil’s regulatory sandbox framework integrates continuous engagement through working groups, fostering a dynamic exchange of ideas between regulators and industry stakeholders. The overarching goal of these reporting mechanisms is to generate knowledge that extends beyond the immediate sandbox participants. Regulatory authorities seek to consolidate best practices, inform future policy developments, and refine regulatory frameworks. For instance, in Rwanda, authorities anticipate that lessons derived from the sandbox experience will directly contribute to the smoother implementation of AI governance mechanisms.

Colombia’s regulatory sandbox has played a pivotal role in advancing the country’s data protection landscape. Notably, it facilitated the issuance of Colombia’s first official guidelines on data protection and AI (External Circular 002, 2024), reinforcing the principle of privacy by design. This regulatory initiative mandates enhanced anonymization techniques and requires detailed documentation of AI system design processes. By embedding these safeguards within the sandbox framework, Colombia’s approach underscores the potential of regulatory sandboxes as instruments for policy innovation, aligning with the country’s statutory framework while fostering responsible AI governance.

Collaboration in effective learning, exit reports, and the requirement of trust. Regulatory sandboxes in jurisdictions such as Mexico and Rwanda prioritize public-private collaboration, fostering new modes of engagement between regulators and industry participants. This approach aligns with contemporary regulatory scholarship, which emphasizes that modern regulatory frameworks increasingly depend on reputation, information-sharing, and sustained interaction with market actors (Fahy, Reference Fahy2022; Ranchordás & Vinci, Reference Ranchordás and Vinci2024). Some regulatory sandboxes, such as the one proposed in Mauritius, explicitly articulate their goal as creating ‘an ideal space to encourage the development and testing of innovative solutions or other systems and collaboration between the public and private sector’ (MPSAR, 2021).

This type of collaboration is also evident in the case of Mexico. The British Embassy in Mexico, in partnership with the Mexican Science and Innovation Network, organized a series of roundtable discussions to share experiences and lessons learned on AI. These discussions aimed to provide an updated overview of the state of AI in Mexico and to foster dialogue among legislators, regulators, and key stakeholders from both countries (la Peña Sissi et al., Reference la Peña Sissi, Ernesto and Cristina2024). This effort led to the development of a regulatory sandbox report, which forms part of a project funded by the Technology Collaboration and Standards Fund of the United Kingdom’s Foreign, Commonwealth & Development Office (FCDO). Through promoting technology standards and international collaboration, particularly in developing countries, this UK fund supports initiatives that leverage technology for development, including projects focused on AI and cybersecurity (UK Government, 2023). FCDO aims to safeguard UK and partner interests in critical technologies, ensure supply chain resilience, address emerging threats, and help position the UK as a global science and technology superpower (Foreign, Commonwealth & Development Office (FCDO), 2025). The effectiveness of these regulatory experiments, therefore, hinges on fostering trust among stakeholders – ensuring that industry actors view regulatory engagement not merely as a compliance obligation but as an opportunity for co-regulatory learning and substantive collaboration.

A critical factor in the success of regulatory sandboxes is access to information, as their effectiveness depends on the availability of relevant data within each experimental setting. Recognizing this requirement, regulatory authorities have established frameworks to collect and present information systematically, often requiring participating entities to submit data that contributes to broader regulatory learning. This necessity is underscored by Brazil’s data protection authority, which has emphasized the importance of fostering algorithmic transparency. Brazil’s General Data Protection Law further reinforces the principle of transparency, mandating that AI systems’ inner workings – including algorithms and data processing techniques – be made visible and explainable.

At the same time, the Brazilian authority acknowledges significant constraints, particularly those related to the protection of intellectual property and trade secrets, which often prevent entities from disclosing critical information. In response, regulatory sandboxes aim to establish mechanisms that facilitate data access while balancing legal and proprietary concerns (ANPD, 2023). Yet, meaningful information exchange within these frameworks ultimately depends on a foundation of trust and mutually agreed-upon principles. Without such trust, the most valuable insights may remain inaccessible, thereby significantly limiting the impact of these experimental environments.

Across multiple jurisdictions, regulatory sandboxes increasingly serve as a mechanism for acquiring AI-related knowledge. In the cases examined, particularly in their initial phases, the sandbox design appears to prioritize this learning function. Only once authorities gain meaningful insights do they consider expanding regulatory experimentation. Furthermore, there is a growing emphasis on producing comprehensive reports that systematically document the learning experiences of regulatory sandboxes. Nearly all of the identified sandboxes emphasize the development of a final exit report, white paper or feedback report to document the information-sharing process and key findings. The reports proposed by authorities in some of these cases serve a dual function: they do not only provide critical insights for regulatory bodies but also contribute to broader policy discussions on AI governance, standard-setting, and supervisory mechanisms. Importantly, these reports extend beyond internal regulatory use, offering valuable resources for external stakeholders – particularly when sandboxes are embedded within national AI or innovation strategies.

However, despite the growing emphasis on documentation and transparency, no final exit reports from these sandboxes have been identified. This may be due to their early stage of implementation or delays in report preparation. Nevertheless, there is evidence of regulatory guidance and decisions emerging from these environments. For instance, in Colombia and Mauritius, authorities have issued guidelines on the responsible use of AI and its adoption in strategic sectors, such as Mauritius’ Chatbot for the Human Resource Management Manual (Circular Letter 20 of 2021). This indicates that regulatory sandboxes are producing tangible outcomes, even if these insights are not always formally documented in comprehensive reports. Constraints related to time and resources likely pose challenges to systematically recording all learnings through official reports and related initiatives.

In sum, based on the analysis of the cases described in Table 1, we have identified elements that highlight learning as the focus of these proposals, rather than the participation of regulators in an innovation process. The envisaged learning processes are also tightly linked to a close collaboration between regulators and stakeholders. The account of regulatory sandboxes described above thus overall suggests that many of the tools and methodologies proposed for the design of these sandboxes will gain from delving deeper into the regulatory learning process rather than focusing exclusively on how they can support innovation.