The Sandbox Principle: Why Some Governments Regulate AI by Experimenting

Most regulation tries to anticipate harm before it happens. But when a technology evolves faster than the evidence about it, prediction becomes guesswork. A cluster of governments has responded with a different instinct: rather than forecast how a new AI system will behave, create a controlled space to watch it behave, and legislate from what is observed. This article examines the agile and experimentalist and facilitating and enabling approaches - the parts of the regulatory menu that treat law as something to be learned.

Introduction

These two approaches sit near the light-touch end of UNESCO's nine-approach spectrum, but they are not passive. They ask the state to do active work: to build spaces for supervised trial, and to cultivate the talent and infrastructure that responsible AI requires (UNESCO, 2026). Both rest on an idea familiar to anyone who has ever learned a skill - that understanding often follows action rather than preceding it.

Regulatory Sandboxes: A Safe-to-Fail Space

The signature instrument of the agile and experimentalist approach is the regulatory sandbox, a scheme first developed in fintech and telecommunications and now extended to AI. The European Union's AI Act defines a sandbox as a controlled framework, established by a competent authority, in which providers may develop, train, validate and test an innovative system under regulatory supervision for a limited time (UNESCO, 2026). Under Article 57, every Member State must ensure such a sandbox is established at national, regional or local level. South Korea's AI Act likewise authorises the government, through Article 19, to support temporary permits and regulatory sandboxes for AI-converged products and services (UNESCO, 2026).

The design converts a controlled experiment into regulatory knowledge. Participants gain limited exceptions and guidance; authorities gain a supervised view of real behaviour and real harm. As the report observes, experimentation can yield new insight into how a proposed framework would actually operate, and the results can inform future regulation - or deregulation (UNESCO, 2026).

Enabling the Conditions: Talent, Infrastructure, Awareness

The facilitating and enabling approach widens the lens from single experiments to the whole environment in which AI is built and used. Japan's AI Act, enacted on 28 May 2025, is a clear example: it seeks to promote research and development (Article 11), facilitate infrastructure sharing (Article 12), develop human talent (Article 14), and advance public education and awareness (Article 15). The last of these directs the national government to promote learning about AI-related technologies and to enhance public understanding through appropriate measures (UNESCO, 2026). South Korea devotes Articles 16 to 23 to support for AI clusters, workforce training and international cooperation (UNESCO, 2026).

This approach reframes the state's role from referee to gardener. Rules still matter, but so do the soil conditions - skills, data infrastructure and public literacy - without which even well-designed rules cannot take root.

The Psychology of Learning by Doing

The sandbox principle mirrors a well-documented feature of how people and organisations learn. Amy Edmondson's research on psychological safety found that teams improve fastest when members can experiment, report failure and surface uncomfortable information without fear of penalty (Edmondson, 1999). A safe-to-fail space does for a regulator what psychological safety does for a team: it makes real information visible while the stakes are still contained. Prediction asks an institution to be right in advance; experimentation lets it be wrong cheaply and correct quickly.

There is a limit worth naming. The report is careful to note that agile arrangements are best suited to situations where risks are low and human rights are not at stake; a sandbox is a tool for uncertainty, not a licence to defer protection where harm is foreseeable (UNESCO, 2026). Used within that boundary, the sandbox principle offers a mature answer to a hard problem - how to govern something you do not yet fully understand. The answer is not to pretend to understand it, but to build the conditions in which understanding can safely accumulate.