A design in which a human must review or approve an automated system's output before it takes effect.
In more detail
Human oversight only works if reviewers have the time, information and authority to disagree; otherwise automation bias turns review into rubber-stamping.