Skill atrophy is what happens to a human capability when AI performs it so reliably that the human stops practicing it. Skills are maintained by use; a delegation that is permanent is also, quietly, a decision to let the underlying skill decay.
The series approaches this through its typology of AI users rather than as an abstract warning. The second essay's "self-automators" outsource their judgment along with their output, and underperform the configurations that keep judgment in human hands. The fourth essay states the counterweight: the tool rewards depth and exposes its absence, meaning the humans who keep their expertise alive get more from the machine, not less, because they can hand it better context and catch its errors.
That points at the design question rather than a ban on delegation. Some skills are safe to let go (few mourn mental long division); others are load-bearing for everything else you do with AI, judgment and domain depth above all. A designed AI×HI relationship distinguishes the two on purpose. Atrophy by drift makes the choice anyway, just without you.