Human oversight of AI is not a control until it is tested
In this opinion, Dr. Mulyadi argues that assigning a person to approve AI output does not prove the person can detect an error, has time to investigate it, or has authority to stop it. Drawing on accounting’s distinction between designed and operating controls, he says boards and audit committees should treat human review as a control-effectiveness question: competent reviewers, realistic capacity, authority to intervene, and independent testing of the full process under realistic conditions, with retesting when technology, staffing, or workload changes. He also warns that automating junior work without replacing learning can spend expertise faster than it develops, and that boards should not accept “a human will check it” as evidence a control works.