Human AI Oversight Often Fails: Approval Rates High, Error Catch Rates as Low as 9%

A technical analysis published on DEV Community challenges the assumption that human-in-the-loop processes reliably improve AI safety. Research on AI coding agents found that while requiring human plan-approval reduced harmful actions from roughly 90% to 60–74%, humans successfully intervened on bad actions only 9–26% of the time across all oversight strategies tested. Two key factors explain this gap: automation bias, where people over-trust system suggestions and reduce scrutiny over time, and the "rubber stamp" pattern, where time-pressured reviewers skim and approve agent proposals without genuine evaluation. The analysis argues that human oversight only provides real safety value when both the consequences are high and a reviewer can realistically detect and correct the problem before harm occurs. Effective oversight must be deliberately engineered — with clear evidence, reversible actions, and adequate review time — rather than assumed from simply placing a human in an approval workflow.