A "human in the loop" isn't just necessary to manage AI's technical failures or in cases of human responsibility. This paper explores the case where AI provides the "correct" answer to the prompt, but the human realizes important factors that weren't even in the prompt. I interpret this as "rubber ducking," but from the POV of automation.