A Contextual-Bandit Oversight Game with Two-Sided Informational Asymmetry

Short summary

Researchers model human oversight of AI agents where both sides hold private information: humans don't know action quality, AI doesn't know reward functions. Using contextual-bandit theory, they characterize the gap between optimal oversight and myopic behavior, identifying where AI can propose harm undetected. The work connects game theory with practical safeguards for autonomous systems.

• •Analyzes human oversight with two-sided private information (asymmetric knowledge)
• •Identifies 'avoidable harm' gap where AI can propose harmful actions undetected
• •Applies contextual-bandit and game theory to improve AI safety mechanisms

Generated with AI, which can make mistakes.