Sandboxes and real-world testing

Lesson 5 of 6 in Governance, Enforcement, Penalties, and Sandboxes.

The Act’s enforcement teeth come paired with an innovation offer. A regulatory sandbox (Arts 57–63) is a controlled framework in which providers develop, train, and test AI systems under the guidance and supervision of the competent authority, against an agreed sandbox plan, for a limited time before market placement. Every Member State had to have at least one operational (alone, jointly, or by joining another State’s) by 2 August 2026.

What does participation actually buy? Three things with legal weight. First, documented regulatory guidance — written proof of the activities performed and an exit report, which providers may use to demonstrate compliance and which authorities must take positively into account in conformity assessment and market surveillance: a head start, not a certificate. Second, a shield: where participants follow the sandbox plan and the authority’s guidance in good faith, authorities impose no administrative fines for infringements of the Act arising in the sandbox. Third, a data basis: Article 59 permits reuse of personal data lawfully collected for other purposes to develop certain public-interest AI systems inside the sandbox — health, environment, energy, public services — under strict safeguards (isolation, deletion, no decisions affecting data subjects). SMEs and startups get priority, free access (Art 62), and microenterprises may even satisfy Article 17’s QMS through a simplified form (Art 63).

The sandbox journey (Arts 57–59)

  1. Apply to the national sandbox

    SMEs and startups get priority access, free of charge (Art 62). Selection criteria are transparent and published.

  2. Agree the sandbox plan with the authority

    Objectives, testing scope, safeguards, duration, exit conditions — the plan is the yardstick for the good-faith fine shield.

  3. Develop, train, test under supervision

    Continuous guidance from the authority; Art 59 data-reuse basis available for qualifying public-interest systems.

  4. Significant risk to health, safety, or fundamental rights emerges?
  5. Mitigate immediately — or the authority suspends/terminates

    The shield never covers pressing ahead through known risk; liability to affected persons remains throughout (Art 57(12)).

  6. Exit report + written proof of activities

    Documents what was tested and what the authority observed.

  7. Report counted positively in conformity assessment and surveillance

    An accelerator for the Art 43 procedure — never a substitute for it.

Outside the sandbox, Articles 60–61 allow testing in real-world conditions of Annex III high-risk systems before market placement — but wrap it in conditions that read like a clinical-trial protocol: an approved real-world testing plan, registration in the EU database, testing limited to six months, extendable once by another six; informed consent from subjects (with a narrow law-enforcement exception where seeking consent would defeat the test, under extra safeguards); protections for vulnerable groups; effective oversight by qualified persons; and — the signature requirement — reversibility: the system’s predictions, recommendations, and decisions must be capable of being reversed or disregarded. Subjects may withdraw at any time without detriment, and any serious incident during testing triggers the Article 73 machinery you already know.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.