We define atrophic delegation as the measurable erosion of human cognitive capability that occurs when AI systems perform tasks in domains where skill compounds with practice. Unlike simple task offloading, atrophic delegation operates through a specific mechanism: the removal of the cognitive engagement required for capability development. We distinguish between two classes of task; those with capped payoff curves, where delegation is safe, and those with uncapped payoff curves, where delegation causes progressive capability loss. We ground this distinction in recent neuroscience evidence demonstrating reduced neural connectivity during LLM-assisted cognition (Kosmyna et al., 2025), the hemispheric asymmetry framework of McGilchrist (2009, 2021), and four decades of automation deskilling research. We argue that atrophic delegation is not a failure mode of poorly designed AI but a structural feature of well-designed AI deployed without a governance architecture that distinguishes between delegable and sovereign tasks.
Kim Lawrance Fischer (Thu,) studied this question.