A large-scale behavioral study published on arXiv has documented how generative AI is altering both the pace and quality of learning. Analyzing a ten-year panel of 3.2 million ALEKS learning interactions, researchers found that students using AI-compatible math problems completed them faster but retained less knowledge, a phenomenon they term cognitive surrender.
The Impact on Study Time
According to the study authored by Rismanchian, Sina; Uzun, Hasan; Matayoshi, Jeffrey; Cosyn, Eric; and Kurd-Misto, Eyad, learning time on AI-susceptible problems—primarily text-based word problems—declined by 2.8% per quarter among college students after ChatGPT's release. This cumulated to a 26.9% decrease over eleven quarters. High schoolers showed an even steeper decline of 31.3%, middle schoolers 9.0%, and Grade 5 students showed no detectable change.
| Student Group | Quarterly Decline | Cumulative Decline (11 quarters) |
|---|---|---|
| College | 2.8% | 26.9% |
| High School | — | 31.3% |
| Middle School | — | 9.0% |
| Grade 5 | — | No detectable change |
The researchers used a quasi-experimental design exploiting variation between tasks more susceptible to AI (text-based) and less susceptible (interactive graph-based problems). This allowed them to isolate the effect of generative AI on study behavior.
Retention and Learning Outcomes
Crucially, the time savings did not come from efficiency gains. Among college students, the post-ChatGPT divergence in study time vanished entirely under proctoring, ruling out broad productivity improvements. Instead, logistic fixed-effects models on randomly assigned proctored retention items revealed a 25% cumulative decline in the odds of a correct response. The same estimator on non-proctored assessment produced a large opposite-signed increase—inconsistent with any platform, cohort, or curriculum explanation.
"These results are among the first large-scale behavioral and outcome evidence that generative AI has altered how students study and the knowledge they build—the population-level indicator of cognitive surrender."
The study addressed a gap between self-report surveys showing little change and small-scale behavioral studies reporting widespread AI use without sufficient duration to measure learning consequences.
Implications for Assessment Policy
The findings have direct implications for educational research, assessment governance, and AI policy. The authors note that the divergence under proctoring—where AI tools are unavailable—shows that students are using AI to bypass learning steps rather than accelerate genuine understanding. This calls into question the validity of unproctored assessments and suggests that educational systems may need to redesign evaluation methods to account for AI's role.
For enterprise technology leaders, the study offers a cautionary parallel: in workforce training and knowledge management systems, deploying generative AI to speed task completion may similarly undermine long-term skill retention. The metric of 'faster completion' alone is insufficient without measuring durable learning outcomes.
The research provides a robust empirical benchmark for understanding how generative AI reshapes cognitive work—whether in classrooms or corporate learning environments.
