There are plenty of examples in the RL training showing it how and when to prompt the human for help or additional information. This is even a common tool in the "plan" mode of many harnesses.

Conversely, it's much harder to represent a lack of doing something

Because it’s just fancy auto-complete.