Attention is the capacity to hold a single difficult thing in mind long enough to get somewhere with it — the precondition for deep reading, sustained argument, and most work worth doing. It is also the resource most obviously under pressure.
A 2026 preprint frames the pressure as a widening asymmetry: as models' effective context windows have expanded enormously, the span of text and thought people sustain unaided appears to have contracted. Its author proposes a self-reinforcing loop, in which each delegation reduces the practice that would maintain the capacity, which lowers the threshold for the next delegation. The mechanism is plausible and, at this stage, better evidenced as a trend than as a proven cause.
This hub gathers the research and reporting on concentration, deep reading, and the slow work that friction protects.
Global reading scores just hit their lowest point on record. The more interesting finding isn't that AI is involved — it's that AI reading aids help almost nobody equally, and the readers they help least are the ones who were already good at reading.
By Joe Jurczyk
The debate has fixated on jobs. The more consequential question is what happens to thinking itself.
By Joe Jurczyk
arXiv preprint • Mar 2026
Argues that as AI systems' effective context windows have grown roughly 4,000-fold since 2017, human attention spans have measurably contracted over the same period, proposing a self-reinforcing "delegation feedback loop" between the two trends.
Societies (MDPI) • Jan 2025
A peer-reviewed survey study linking frequent AI tool reliance to lower critical thinking scores, especially among users who never built the underlying skill unassisted.