The Collapse of Supervision - V1.0

This paper develops a dynamic theory of human–AI supervision in high-reliability systems. Building on the human–AI contracting framework of Bastani and Cachon (2025), it introduces time and attention as endogenous state variables governing the feasibility of oversight. The core result is that supervision collapses dynamically even under incentive-compatible contracts. As AI systems become increasingly reliable, failures occur too infrequently to replenish human attention, causing the effective cost of inspection to grow endogenously over time. Monitoring therefore persists in form but not in substance. The analysis further shows that holding mean reliability constant, clustered failures can sustain supervision by restoring informational salience. This result reframes the apparent preference for imperfect systems as a temporal coordination mechanism rather than a behavioral anomaly. More broadly, the paper characterizes institutional failure as a temporal mismatch between continuous machine operation, linear institutional time, and rhythmic human cognition. The findings have implications for AI governance, organizational design, and contract theory, suggesting that effective oversight requires temporal structuring rather than stronger incentives alone.