Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build

Rismanchian, Uzun, Matayoshi, Cosyn & Kurd-Misto (2026) — ALEKS/McGraw Hill. arXiv preprint (cs.CY, cs.AI, cs.HC).

📄 Full text (arXiv) — updated to v3 (revised 13 Jul 2026; raw file refreshed from the v3 PDF)

Summary

This landmark study provides the first large-scale behavioral and outcome evidence that generative-ai has fundamentally altered how students study and what they retain. Using a ten-year panel of 3.2 million ALEKS learning interactions and complementary placement-assessment data, the authors employ a quasi-experimental design that exploits within-curriculum variation in AI susceptibility: text-based word problems that can be transcribed into AI prompts serve as the treated group, while graph-based problems requiring interactive platform manipulation serve as the comparison.

Key Findings

Time-on-Task Decline

Learning time on AI-susceptible problems declined at a rate of 2.8% per quarter among college students after ChatGPT's release, cumulating to a 26.9% reduction over eleven quarters. The effect varied by age group:

High school: 31.3% cumulative decline

Middle school: 9.0% cumulative decline

Grade 5: No detectable change

Proctoring Eliminates the Effect

The divergence in study time vanishes entirely under proctoring for college students, ruling out general efficiency improvements as an explanation. This strongly implicates off-platform AI use rather than any platform, cohort, or curriculum change.

Learning Outcome Impact

Proctored retention items: 25% cumulative decline in odds of correct response

Non-proctored assessment: Large opposite-signed increase — impossible to attribute to anything other than AI assistance

The "Cognitive Surrender" Concept

The authors introduce the term cognitive surrender to describe students offloading thinking to generative AI, producing a measurable population-level decline in durable knowledge. This represents a fundamental shift from cognitive-offloading as a deliberate metacognitive strategy to a passive, unreflective dependency.

Implications for the Wiki

This study provides the strongest empirical evidence yet for the over-reliance phenomenon documented across the wiki. It extends findings from genai-performance-vs-learning and cognitive-shift-ai-education by demonstrating effects at population scale with objective behavioral measures rather than self-report. The findings have direct implications for:

academic-integrity policy and assessment governance

ai-literacy curriculum design

learning-gains measurement methodology

Educational regulation and AI policy

over-reliance — The phenomenon of AI dependency this study quantifies at scale

cognitive-offloading — Related but distinct; cognitive surrender is passive and unreflective

cognitive-shift-ai-education — Complementary evidence of AI-driven cognitive change in students

genai-performance-vs-learning — Distinguishing performance gains from actual learning

generative-ai-guardrails-harm-learning — Complementary causal RCT: unguarded GPT-4 tutor raised practice +48% but cut unassisted exam scores 17%; guardrails neutralized the harm

learning-gains — The measurement framework this study challenges

student-experience — Student behavior and outcomes

academic-integrity — Assessment implications of AI use

stem-education — Math education context

genai-availability-grades-satisfaction — Generative AI Availability, Grades, and Student Satisfaction

Citation

APA: Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026). Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build. arXiv:2605.21629.