Real-world testing and regulatory sandboxes (Articles 57-60)

Short answer: The AI Act wants to give innovation room without giving up protection. It does so through two testing routes: the regulatory sandbox, a controlled environment under the supervision of the competent authority (Articles 57-59), and testing in real-world conditions outside the sandbox (Article 60). Both require a plan, safeguards and — where people are involved — informed consent.

The regulatory sandbox (Articles 57-59)

Each Member State sets up at least one AI regulatory sandbox (alone or shared) within two years of entry into force. In the sandbox, providers can develop, train, test and validate an innovative AI system under guidance before market entry, within an agreed plan and a limited period.

The authority gives advice and guidance and can check compliance with the regulation. Participation does not exempt the provider from its high-risk obligations, but it shields against fines for infringements that occur within the sandbox plan and in good faith. Personal data may, under conditions, be processed to develop AI in the public interest.

Testing in real-world conditions (Article 60)

Sometimes a system must be tried outside the lab. Article 60 allows a high-risk system to be tested in real-world conditions outside a sandbox, provided there is an approved testing plan and strict conditions are met:

• the test is registered and has a defined, in principle limited duration;
• there are safeguards against risks to health, safety and fundamental rights;
• decisions with legal effects can be reversed or disregarded;
• the provider can immediately stop the test and erase the data.

This is expressly not a placing on the market: the system has not yet been declared conform.

Protecting test subjects

Where natural persons are involved in the test, informed consent applies: the subject knows they are taking part, understands what the test involves, and can withdraw without detriment. The provider keeps the consent and makes the relevant documentation available to the authority.

Why this matters in the chain

The sandbox and real-world test are the phase before the roles of importer and distributor: only after a successful conformity assessment does the system truly reach the market. The testing routes lower the threshold for making innovation demonstrably safe.

What to do

• Choose the right route: sandbox for guidance and legal certainty, real-world test for field validation.
• Write a testing plan covering duration, risks, safeguards and a stop-and-erase procedure.
• Arrange consent where people are involved and keep the evidence.
• Notify and register the test with the competent authority.
• Anchor it in governance: include testing routes in your AI governance framework.

The testing rules are the AI Act's invitation to innovate in full view of the regulator — not around it.