Why it matters
The researchers analyzed monthly closed-book exams, entrance exams, homework scores, and time spent on homework across nine subjects and six grade levels. Their estimates show a widening gap between visible assignment performance and retained knowledge: homework improved quickly, while exam declines emerged within months and grew with longer exposure. The largest losses appeared in social sciences, followed by STEM and languages, and were especially pronounced for younger students, high achievers, and boys.
The result is not an argument that every use of AI harms learning. Students who kept spending about as much time on homework as non-users experienced comparatively small losses. The governance problem is therefore one of incentives and measurement: if schools reward answer quality without testing the reasoning behind it, AI can make academic performance look stronger while the underlying capability becomes weaker.
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
CEPR — The Generative AI Learning Penalty
