Abstract
Nearly every AI model now warns its user, in some form, that it can be wrong. This note treats that warning as a symptom of a third, most consequential level of intellectual delegation, in which a person hands over not just thinking but checking to a machine while remaining confident in their own intelligence. A short argument first establishes why modest usefulness costs strictly more than grand uselessness, in both entropy and effort. A coverage bound then shows the checked fraction of AI-assisted output collapsing toward zero as volume grows, independent of diligence (Vuong, 2026a, Prop. 3). A cubic captureand-hysteresis model, completing a hypothesis a companion note left explicitly tentative, shows lapsed vigilance is harder to recover than it was to lose. A proficiency/effectiveness decoupling result then shows visible polish and genuine understanding drifting apart without anyone deciding this should happen. A closing corollary identifies the resulting absorbing state exactly with the grand uselessness established at the outset, and Kingfisher-style "relearning to be stupid" with the activation cost of modest usefulness. None of the results is new; the contribution, if any, is the mapping.

This publication has 2 references indexed in Scilit: