Human oversight and authority
contrôle humain, article 26(2)
The deployer assigns human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support. In practice: identifying specific individuals, training them on the specific system rather than on AI in general, and giving them real authority to override, suspend or discontinue the output.
This applies to high-risk systems, and unlike Article 4 it sits within the harmonised penalty scale.
What it means for an SME running AI
A sign-off with no power to stop is not a sign-off. If nobody can name who may halt the chain, the requirement is not met, whatever the procedure says.
This is what the survey measures See how it is measured →
Or answer six questions on your own uses Take the check →