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

Yunjin Tong

We study runtime human oversight of an AI agent when private information runs in both directions: the human privately knows her reward function, while the AI privately knows the qu…