Desirable difficulties is a learning-science term for effort that improves long-term retention (because it forces deeper processing).

If AI can draft, iterate, and error-check successfully most of the time (e.g., ~78% in some evaluations), then doing that work yourself is often inefficient if your goal is speed and output quality.

So the question is what should we still do manually to build the subject expertise needed to (1) specify requirements clearly and (2) evaluate AI outputs reliably?