Article 60: Testing of High-Risk AI Systems in Real-World Conditions Outside AI Regulatory Sandboxes

Article 60 recognises that some high-risk AI systems cannot be adequately tested in sandboxes or laboratory environments alone — real-world complexity may be essential to validating safety and performance. The article creates a structured pathway for such testing with the following requirements:

1. Real-world testing plan — the provider must prepare a detailed plan and submit it to the relevant market surveillance authority for approval before testing begins. 2. Informed consent — all persons who are subjects of the testing must give informed consent in accordance with Article 61. Consent must be freely given, specific, informed, and unambiguous. 3. Monitoring and halt capability — the provider must continuously monitor outcomes during testing and must be able to immediately halt the test if a serious incident occurs or if risks to health, safety, or fundamental rights materialise. 4. Time and scale limitations — testing must be limited in duration and scope to what is necessary for the validation objectives. 5. Safeguards for test subjects — appropriate measures must protect the rights and safety of affected persons throughout the test.

The plan must also address what happens to data collected during testing and how results are incorporated into the conformity assessment process.

Article 60 is a complement to the sandbox framework, not a replacement:

• Sandboxes (Article 57) offer a supervised environment within which authorities guide development. They are best suited for early-stage testing and compliance preparation.
• Real-world testing under Article 60 addresses later-stage validation where the AI system needs exposure to live conditions to demonstrate safety and efficacy.

Importantly, real-world testing under Article 60 does not constitute placing on the market or putting into service. It is a pre-market activity. However, the results are expected to feed directly into the provider's risk management system (Article 9) and technical documentation (Article 11).

• Article 57 — sandbox framework (complementary pathway for earlier-stage testing).
• Article 61 — informed consent requirements for real-world testing participants.
• Article 73 — serious incident reporting obligations (applicable if incidents occur during testing).
• Article 76 — confidentiality obligations for authorities overseeing testing.
• Article 9 — risk management system (real-world test results feed back into risk analysis).
• Article 113 — application dates.

1. Start with the testing plan — define clear validation objectives, the specific real-world conditions needed, duration, geographic scope, number of subjects, and the metrics you will measure. 2. Engage the market surveillance authority early — submit your plan well in advance and be prepared for iterative review. Authorities may impose additional conditions. 3. Build informed consent into your design — work with Article 61 requirements from the start. Consent forms should explain the AI system, the testing conditions, risks, and the right to withdraw. 4. Implement a kill switch — you must be able to halt the test immediately at any point. Design technical and operational procedures for this before testing begins. 5. Document everything — test results, incidents, consent records, and authority correspondence all become part of your conformity evidence.