The skill of checking the machine
Human oversight means little unless someone has the knowledge, time, and authority to disagree.
A workflow ends with a box labelled human review. That label sounds reassuring, but the person reviewing may be rushed, unfamiliar with the subject, or discouraged from changing an automated recommendation. A human presence is not automatically meaningful oversight.
A human review box is reassuring only if the reviewer can understand and change the result. Time, expertise, and permission to disagree are the conditions that make oversight real.
Checking requires knowing what a correct result looks like and where the system is likely to fail. It also requires a manageable volume of work. If a reviewer is asked to inspect hundreds of complex outputs in minutes, the process may offer the appearance of accountability without its conditions.
Design review around consequences. A low-stakes draft may need a quick edit. A decision affecting money, rights, health, or access deserves stronger evidence and a route to appeal. The person affected should not be expected to discover an invisible error alone.
Checking also needs a manageable volume. An organisation cannot reasonably demand careful judgment while arranging the workload so that approval is the only action anyone has time to perform.
Workers should have explicit permission to stop the process, request more information, and record disagreement. Training can include deliberately flawed outputs so reviewers practice detecting mistakes rather than merely approving plausible ones.
The question is not whether a human touched the final answer. It is whether that person could understand, challenge, and change it. A system becomes more accountable when oversight is treated as skilled work with resources, not as a ceremonial signature at the end.