What you hand to the model you stop practising. Delegate what you already understand; learn yourself what you will need to judge.
In a field experiment with high-school maths students (Bastani et al., PNAS 2025), unrestricted access to a chat assistant raised scores during practice and lowered them once access was removed, below students who never had it; a tutor that gave hints instead of answers largely removed the harm. The checks in the other cards here — finding the load-bearing claim, spotting a missing explanation — need the very skill that unchecked delegation erodes. One experiment in one setting is not a law of nature, but it agrees with an older result.
Lisanne Bainbridge, “Ironies of Automation” (Automatica, 1983), studied process-control operators: automation takes over the routine work and leaves the human to monitor and to step in when it fails. But monitoring a system that rarely fails is something people do badly, and the skill needed to take over decays exactly because the automation does the work. The more reliable the automation, the less practised the person who must handle its failures. A code assistant inverts nothing here: the engineer who no longer writes SQL by hand is the one asked to judge whether a generated query will lock a table.
How to delegate without the atrophy. Separate the work you understand and want done from the work you need to keep judging. Hand over the first freely — boilerplate, a migration you could write in your sleep. For the second, ask for hints, critiques or a plan instead of the finished answer, write the first version yourself and compare, and periodically do the task unaided to check that the skill is still there. The cost is speed now; the trade is against review capacity later (59). Chesterton’s Fence, both ways (61) covers code without an author who understood it; this card covers the author who stops understanding.