Human oversight of AI is the ability of people with appropriate authority and knowledge to understand, monitor and intervene in an AI system’s use. Effective oversight gives a human a meaningful opportunity to question or change an outcome, rather than asking them to approve a decision they cannot assess.

Two common descriptions are human-in-the-loop and human-on-the-loop. In a human-in-the-loop arrangement, the system pauses for human input or approval before a defined action is taken. In a human-on-the-loop arrangement, the system acts within set limits while a person monitors its operation and can intervene. Neither label guarantees good oversight: the details of authority, timing, information and workload determine whether a person can make a difference.

Make intervention practical

People need to know what the system is for, what it cannot reliably do, which cases need attention and what evidence supports an output. They need enough time to review that information and a way to challenge or override the system. If the interface hides uncertainty, the queue is too large or management rewards speed alone, a nominal approval step can become rubber-stamping.

Oversight design should match the impact and reversibility of the action. A low-impact suggestion may be checked through sampling and escalation rules. A decision that can deny a person an opportunity or affect safety may require individual review, reasons that can be examined, and a route to reconsider the result. The organisation should test whether those controls work under realistic time pressure and when the system is wrong.

The named human

Every deployed AI system should have a named human owner. The owner is accountable for its approved purpose and operation, knows which teams perform testing and monitoring, and has authority to request a review or pause its use. This role cannot be assigned vaguely to a committee. Committees may set policy and challenge evidence, but a specific person needs to own each system and act when a threshold is crossed.

That owner should be supported by technical specialists, legal and risk teams, frontline users and affected people where appropriate. Responsibilities should be documented across the lifecycle: who approves deployment, who monitors performance, who handles incidents and who decides whether the system may resume after a pause. The owner also needs a named deputy or continuity plan for absence.

Human oversight is not a substitute for reliable design or appropriate limits. People can make mistakes, become over-reliant on automation or lack the context to recognise an error. Good governance combines human judgement with system testing, clear boundaries, records, escalation routes and authority to intervene. The aim is to make responsibility real and keep decisions contestable when the consequences matter.