heygrc
Glossary

Human oversight

Human oversight, a requirement of the EU AI Act (Art. 14) for high-risk AI systems, is the expectation that a person can effectively oversee the system, including monitoring it, interpreting its output, and being able to disregard, override, or reverse that output or stop the system, proportionate to the risk and context.

In code

It is enforced by the path that routes a decision to a person. It weakens when a change auto-confirms a model's output (for example above a confidence threshold) and removes the step where a human could review or override it, with no other effective oversight in place.