Human oversight means that natural persons can effectively monitor an AI system while it is in use and intervene when needed. Art. 14 of the AI Act requires this for high-risk AI systems. The overseers understand the capabilities and limits of the system, interpret its outputs correctly, can disregard or override a result and can halt operation through a stop function. For remote biometric identification, at least two persons must confirm an identification before any action is taken on it (Art. 14(5)).
The duty is split between two roles. The provider designs the system so that oversight is possible. The deployer assigns persons with the necessary competence, training and authority and gives them the support they need (Art. 26). Oversight loses its effect once approvals become routine, so automation bias belongs in every training for overseers. The risks of bypassed oversight are shown in MU-05 Circumventing human oversight.