Six papers published this week, read separately, each describe a local finding about AI performance. Read together, they describe a single structural problem: AI tools extract human capability as a operating cost, not a one-time price. The productivity gains from AI adoption may be real, but they are partly borrowed from the future competence of the people using the tools.
These papers converge on an accounting problem. AI productivity gains are measured at the point of output — tasks completed, code written, time saved. The liabilities accumulate somewhere else: in skill atrophy that shows up only when the tool is absent, in integration failures that emerge weeks after the commit, in benchmark scores that overstate real-world performance by 20 points. The structural driver is a mismatch between measurement time and consequence time. What remains unknown is whether the atrophy is reversible — whether removing the tool restores the skill, or whether the loss is permanent as the theoretical paper argues.
Watch for any enterprise software team publishing before-and-after data on developer debugging time or code review duration after six or more months of mandatory AI coding agent use.