Human Oversight (Human-in-the-Loop) (HITL)

Designs differ in how closely the person is involved. In human-in-the-loop, a person approves each output before it takes effect. In human-on-the-loop, the system acts alone but a person watches and can intervene. In human-in-command, people decide whether and how the system is used at all. Oversight only works if the reviewer has the authority, time, skill and information to disagree. A reviewer who sees hundreds of cases an hour and almost always clicks accept offers little real control, a problem called automation bias. The EU AI Act requires effective human oversight for high-risk systems, and data protection laws such as the GDPR limit purely automated decisions with significant effects. Firms document who oversees what, how overrides are recorded and how reviewers are trained. A reviewer with real power to disagree catches errors the model cannot see in itself, while a rubber stamp adds only delay.

A fraud model flags card transactions for blocking. Low-score cases are released, but blocks above a set value go to an analyst, who sees the main reasons for the alert and can override it. The card issuer tracks how often analysts overturn the model. When the override rate drops near zero, the team samples cases to test whether analysts are really checking or simply accepting.