Human oversight of AI means that a human being with sufficient competence and information decides on, or can overturn, the output of a system that has an impact on people. Formally rubber-stamping proposals is not oversight – if the reviewer has neither the time, the underlying materials, nor the authority to decide otherwise, this amounts to automated decision-making with the appearance of control. Functional oversight therefore requires several things at once: the person knows the system's limitations and when it is wrong, sees the reasons behind the proposal and not just the result, has a genuine possibility of deciding otherwise without penalty, and has adequate time to do so. This must be supplemented with a record of who decided and when. For sensitive processes such as recruitment or client assessment, this is precisely the first thing a check examines, and it is also what protects the company most in a dispute.
See also: High-risk AI system, Fundamental rights impact assessment, Technical documentation of an AI system.
