> Questions are "good questions" because they are good questions to humans.

This is what reinforcement learning human feedback (RLHF) is for.