What is human oversight of AI?

What human oversight requires in practice

Genuine human oversight of an AI system requires three things: capability, authority, and genuine intervention power. Capability means the human responsible for oversight actually understands enough about the AI system to identify when it is behaving unexpectedly — they cannot exercise meaningful oversight if they treat the system as a black box and accept all its outputs uncritically. Authority means the overseer has the organisational power to stop, modify, or override the AI system when they determine it is necessary — nominal oversight by someone who cannot actually change anything is not oversight. Genuine intervention power means the technical and procedural mechanisms to pause or modify the system are actually in place and usable. For many organisations, the gap between nominally having human oversight and genuinely having it is significant — particularly where AI has been integrated into workflows without adequate review of who can actually intervene when something goes wrong.

EU AI Act human oversight obligations

For high-risk AI systems, the EU AI Act mandates human oversight as a non-negotiable requirement. The Act specifies that high-risk AI systems must be designed and developed to allow for effective human oversight during their use. They must include interface features that allow those responsible for oversight to monitor performance, detect anomalies, and intervene. The individuals responsible for oversight must have the necessary competence, training, and authority to exercise it properly — which is where Article 4's AI literacy obligation connects directly to human oversight. Deployers of high-risk AI must assign human oversight to specific individuals, ensure those individuals are trained, and create the organisational conditions for genuine oversight to operate. A human name on a governance chart is not enough — the EU AI Act requires that human oversight is real, documented, and demonstrable.