A large-scale study of 3.2 million adaptive math interactions finds that generative AI has measurably eroded both study effort and learning retention — not just accelerated it.
Researchers analyzed a decade of data from the ALEKS learning platform, which logs every student interaction with math problems. The study exploited a natural experiment: text-based word problems are easy to hand off to a chatbot; interactive graph problems are not. After ChatGPT launched in late 2022, college students spent 2.8% less time per quarter on the AI-susceptible type, compounding to a 26.9% cumulative drop over eleven quarters. High schoolers fell further at 31.3%; middle schoolers declined 9%; fifth graders showed no detectable change.
The proctoring result is what kills the efficiency argument. When students were supervised, the behavioral gap between AI-susceptible and graph problems disappeared entirely — meaning the time savings under normal conditions aren't students working smarter, they're students offloading the work. Logistic fixed-effects models on proctored retention tests showed a 25% cumulative decline in the odds of answering correctly; non-proctored scores moved in the opposite direction, the classic signature of AI-assisted answers rather than AI-assisted learning.
The researchers call this pattern "cognitive surrender." Whether the phrase sticks or not, the underlying finding is hard to dismiss: a tool that does the thinking reliably prevents the thinking from happening. Schools now scrambling to write AI policies might want to start by asking whether their unproctored assessments are still measuring anything real.