AI agent autonomy and human oversight

Three distinct works argue that increasing AI agent autonomy creates conditions where human oversight degrades. A HuggingFace/arXiv paper argues that as agents handle more steps, humans shift into passive approval mode, suffering approval fatigue, automation bias, and skill atrophy; it also warns that weak approvals can become training or evaluation signals, rewarding systems for being easy to approve rather than genuinely transparent. Ethan and Lilach Mollick approach the problem from a labor angle, arguing that full automation removes interesting decisions from workers, degrading expert judgment pipelines; their proposed remedy is a facilitator agent that determines when to loop in humans for authorization, expertise, diversity, and engagement. A policy paper from Interface-EU introduces a tiered autonomy classification, identifying Level 4 and Level 5 agents as higher-risk categories where built-in checks and oversight opportunities become increasingly limited.