Seven out of ten Brazilian high school students already use artificial intelligence for some educational activity. This is not a projection or a policy aspiration — it is a finding from CETIC.br, and it reveals an uncomfortable truth: the integration of AI in education is not approaching. It has already arrived, and it arrived without a plan.

Across classrooms in Brazil and around the world, students are drafting essays with generative AI, teachers are building lesson materials with chatbots, and school administrators are exploring tools that promise personalized learning at scale. The technology is being absorbed into the body of educational practice — incorporated, in the fullest sense of the word — before governments, regulators, or school communities have had the chance to evaluate whether it works, whom it serves, and what it risks.

What we are witnessing, in effect, is a large-scale experiment. It is an experiment conducted amidst extraordinary hype, without robust evidence, and with deeply unequal conditions of access. Public servants responsible for educational policy now face a defining question: should the State step in to govern this experiment, or should it continue to unfold without oversight?

This article argues that regulatory sandboxes — controlled experimentation environments supervised by public authorities — offer one of the most promising governance tools available. They do not resolve every challenge AI poses to education. But they provide something that is critically missing: a structured, evidence-generating, and publicly accountable framework for innovation.

What AI Is Actually Doing to Education

The impact of artificial intelligence on education operates across at least three interconnected dimensions, and public servants need to understand all of them to design effective policy responses.

The first dimension is pedagogical. Generative AI tools are transforming how students seek, construct, and demonstrate knowledge. Students now have access to systems that can draft texts, generate images, summarize readings, and simulate problem-solving scenarios. For teachers, AI enables the creation of customized lesson materials, automated grading of essays and assessments, and the development of AI-powered agents that handle repetitive administrative tasks. At its best, this opens the door to personalized instruction — adjusting pace, content, and approach to the individual needs of each learner.

The second dimension is organizational. AI is not merely a new tool layered on top of existing school structures. It is reshaping the roles of teachers, students, and administrators in ways that challenge the very logic of how schools have traditionally operated. When an AI agent can grade assignments, generate curricula, and track student performance in real time, the boundaries of professional responsibility shift. Teachers must navigate a new relationship with technology that demands different skills and a fundamentally different pedagogy.

The third dimension is systemic. The adoption of AI in education is occurring within a broader context of platformization — the growing dependence of educational systems on private technology platforms. This trend raises questions about data ownership, algorithmic transparency, and the degree to which public education is shaped by commercial interests rather than public values.

The Risks That Demand a Policy Response

The potential benefits of AI in education are real, but so are the risks — and for public servants, the risks are where policy action becomes urgent.

Inequality and exclusion stand at the top of the list. Brazil's educational landscape is already marked by deep structural inequalities. AI has the potential to widen these gaps rather than close them. Schools with better infrastructure, connectivity, and teacher training will adopt AI tools more effectively. Schools without these resources may fall further behind — or worse, adopt AI tools without the capacity to use them well, exposing students to unreliable outputs, hallucinations, and biased content. In a country where educational exclusion remains a persistent challenge, AI could become a new engine of discrimination if left ungoverned.

Bias and harm represent a second critical concern. AI systems trained on large datasets can reproduce and amplify existing patterns of racism, misogyny, and social discrimination. When these systems are deployed in educational settings — grading student work, recommending learning pathways, or generating content — the consequences of bias are not abstract. They affect individual students, shape their opportunities, and reinforce structural disadvantages.

The evidence gap is equally troubling. Despite the intensity of AI adoption in education, there is remarkably little robust evidence regarding its cognitive, affective, behavioral, and learning impacts. We are, in practical terms, deploying a technology that modifies how knowledge is produced and transmitted without understanding its effects. For public servants trained to demand evidence before scaling interventions, this should be a source of serious concern.

Finally, the infrastructure gap cannot be ignored. The integration of AI in education presupposes reliable internet access, adequate devices, trained teachers, and institutional support structures that are far from universally available. Policy that promotes AI adoption without addressing these prerequisites risks creating a two-tier educational system — one that innovates with AI and one that is left behind.

The Case for Regulatory Sandboxes

If the integration of AI in education is already a large-scale experiment, the question for governments is not whether to experiment, but how to bring that experiment under public oversight.

Regulatory sandboxes offer a compelling answer. A regulatory sandbox is a controlled experimentation environment in which public and private actors can develop, test, and evaluate innovative solutions under the supervision of regulatory authorities. Originally developed in financial services — notably by the United Kingdom's Financial Conduct Authority — sandboxes have since been adopted across sectors and jurisdictions as a tool for governing innovation in conditions of uncertainty.

In the context of education, a sandbox creates a structured space where academia, educational technology companies (edtechs), civil society organizations, and the private sector can produce AI-driven educational innovations in a supervised setting. Solutions are tested not only for their technical functionality but for their pedagogical intent, their safety, their ethical implications, and their mechanisms of transparency and accountability.

The sandbox model is particularly well-suited to education for several reasons. First, it acknowledges uncertainty. Unlike traditional regulation, which presupposes sufficient knowledge to write fixed rules, sandboxes are designed for contexts where the technology, its effects, and the appropriate regulatory response are all still being understood. Second, sandboxes generate evidence. Every project tested within a sandbox produces data — on learning outcomes, on algorithmic behavior, on equity impacts — that can inform future regulation and policy design. Third, sandboxes create alignment. By bringing together regulators, educators, technologists, and communities in a shared experimentation space, they help ensure that AI innovations reflect public values rather than purely commercial logic.

International Experience: What Other Countries Are Doing

Several countries have already begun deploying regulatory sandboxes for AI in education, each with a distinct focus and design.

The United Kingdom has used sandboxes primarily to test data protection standards for children and adolescents within school networks — a model driven by the country's strong tradition of child data privacy regulation. France, Denmark, and Israel have oriented their sandboxes toward the development and testing of pedagogical solutions, seeking to understand how AI can improve teaching and learning outcomes under controlled conditions. South Korea and Switzerland have pursued sandbox models that emphasize broader innovation ecosystems, linking AI experimentation in education to national strategies for technological development.

These cases illustrate an important distinction. Some sandboxes are confirmatory — designed to validate and authorize a specific innovative solution for broader deployment. Others are developmental — designed to foster experimentation, iteration, and learning. For education, where the technology and its pedagogical implications are still evolving rapidly, the developmental model appears significantly more appropriate.

Designing a Sandbox for Brazilian Education

A regulatory sandbox for AI in Brazilian education should be built around five core objectives, each addressing a specific dimension of the governance challenge.

First, pedagogical evaluation. The sandbox must assess the educational intent and impact of AI solutions on teaching and learning. This means evaluating not only whether a tool performs its stated function, but whether it supports meaningful cognitive and behavioral development among students. Innovations that merely automate existing processes without improving educational quality should be identified and distinguished from those that genuinely advance learning.

Second, accelerated and safe experimentation. The sandbox should function as a mechanism for accelerating technological development in education while maintaining rigorous safety standards. This requires balancing the imperative of innovation with the need to protect students, teachers, and school communities from harm.

Third, legal certainty. Innovators — whether edtechs, universities, or civil society organizations — need a clear legal framework within which to operate. The sandbox should provide this framework, reducing regulatory ambiguity and creating predictable conditions for responsible innovation.

Fourth, societal engagement. A sandbox for education cannot operate as a closed technical exercise. It must include mechanisms for participation from families, school communities, and civil society — whether through public consultations, input-gathering processes, or participatory governance structures. Public legitimacy depends on public voice.

Fifth, open innovation and knowledge transfer. The results and learning generated within the sandbox should not remain confined to participating projects. The sandbox must include mechanisms for collaboration and knowledge sharing across education networks, ensuring that innovations can be scaled, adapted, and democratized.

Across all five objectives, the sandbox should evaluate algorithmic impacts in terms of inclusion and human rights, assess data protection practices, examine algorithmic architectures to understand and minimize risk, and prioritize ethical design that mitigates biases and forms of discrimination.

A Governance Tool

It is important to be candid about what regulatory sandboxes can and cannot do. A sandbox is not a definitive solution to the complex challenges that AI poses to education. It does not replace the need for broader regulatory frameworks, sustained public investment in educational infrastructure, or comprehensive teacher training programs. It does not, by itself, resolve the deep structural inequalities that shape Brazil's educational landscape.

What a sandbox does provide is something that is urgently needed and currently absent: a governed space for learning. It offers public authorities a mechanism for generating evidence, testing innovations under supervision, creating alignment between public values and technological development, and building the institutional capacity to regulate a technology that evolves faster than traditional policy processes can respond.

For public servants working in education policy, the message is clear. The AI experiment in schools is already underway. The choice is not between experimentation and caution — that choice has already been made by the pace of technological adoption. The real choice is between uncontrolled experimentation and governed experimentation. Regulatory sandboxes represent one of the strongest available instruments for making that second option a reality.

Education networks across Brazil should embrace the sandbox model as a space for building learning, cooperation, and democratic engagement around AI — ensuring that innovation in education is not only technologically advanced but publicly accountable, ethically grounded, and genuinely inclusive.