Human oversight is the principle of the EU AI Act (AI Act) according to which AI systems must be designed so that trained individuals can understand what they do, monitor their operation, and intervene when necessary. Article 14 explicitly requires this for high-risk systems, and the spirit extends to any company that deploys agents to act on customers, money, or data.
What the supervisor must be able to do
- Understand the system's capabilities and limitations, and detect anomalies or unexpected results.
- Avoid relying automatically on what the machine proposes (the regulation refers to avoiding 'automation bias').
- Correctly interpret the output and decide whether to use, correct, or discard it.
- Intervene: stop the system, undo an action, or prevent it from executing.
Training
Oversight requires knowledge. That is why Article 4 of the same regulation obliges companies to ensure the AI literacy of the people operating these systems: you cannot oversee what you do not understand.
Relationship with HITL
Human in the loop (HITL) is the practical way to comply: placing a person's approval within the flow of critical actions, with minimum permissions and explicit limits for each agent. The question that summarises both is the same: who directs the system, with what permissions, and with what limits.