Answers are grounded in this paper and cite their sources.
Hidden capture is a loss of skill that the people losing it do not notice, combined with a dependence that does not reverse when the pressure behind it eases. The Correction Freeze Problem predicts this pattern, but testing it on human thinking would take decades of data that do not yet exist for generative AI. We therefore built a computer simulation of a population, calibrated to what can be measured today, and followed it across generations in two worlds that share the same machines. In World M, AI mostly replaces people’s own practice, and each generation learns what normal ability looks like from the one before. In World N, AI mostly supports practice, and each generation’s sense of normal ability is anchored in direct experience, including time in nature. The model reproduces today’s rise of AI-written text in student work, the share of people who check AI output, and how loosely people’s opinions of their own thinking track tested performance. The simulations show what a captured world would be like from the inside. Delegation stays modest for a while and then sweeps through most of the population within a few years, because each person who hands over their thinking makes it more normal for everyone else. Skill then drains away without being felt. In World M, the average person’s unaided ability fell to about the bottom tenth of the population before AI, yet people registered only a small part of that loss, about a quarter even a century later, because their machine-assisted work looked as good as ever and everyone around them had declined too. The habit of checking disappeared first: at full delegation, only two people in a hundred still compared the machine’s answers with their own judgment, and when the machine’s error rate was quietly doubled, about one in a hundred noticed within five years. In an illustrative century, easing the pressure after a generation of heavy delegation restored only World N; the other two worlds were still captured half a century later. Staying out of the trap proved far easier than climbing out of it. These findings are conditional projections, not forecasts. Across 2,176 sensitivity runs, the unseen loss appeared every time, and its size depended mainly on two things societies can influence now: whether AI replaces or supports people’s own practice, and whether each generation anchors its sense of competence in direct experience. Whether capture can be undone depended on one quantity not yet measured, how much an individual’s reliance on AI changes from year to year; if reliance churns, capture becomes reversible, though it stays hidden. Several earlier results weakened as the model was corrected, and the main pre-specified test of a final extension came back negative; both are reported.
No takes yet. Share an insight, caveat, or question.
Vuong et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: