Modeling Human Beliefs about AI Behavior for Scalable Oversight