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

Synthesis: Rismanchian, Uzun, Matayoshi, Cosyn, and Kurd-Misto (2026) provide 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 plus 12.2 million ALEKS PPL placement-assessment response times, a quasi-experimental design exploits within-curriculum variation in AI susceptibility: text-based word problems (transcribable into AI prompts) are treated, while interactive graph-based problems (requiring platform manipulation) serve as control. Learning time on AI-susceptible problems fell 2.8% per quarter among college students after ChatGPT's release (26.9% cumulative over eleven quarters), the divergence vanishes entirely under proctoring, and proctored retention items show a 25% cumulative decline in odds of correct response. The authors call this population-level displacement of thinking Cognitive Surrender — a shift from deliberate offloading to passive dependency.

Study design and scale

The study is notable for its scale and its quasi-experimental identification strategy:

• Time-on-task: a ten-year panel of 3.2 million ALEKS learning interactions (a mastery-based math learning platform).
• Learning outcomes and proctoring: 12.2 million ALEKS PPL placement-assessment response times, used to test whether the time decline persists when AI use is blocked (proctored) versus when it is not.
• Identification: within-curriculum variation in AI susceptibility — text-based word problems that can be transcribed into prompts for an Large Language Models (LLMs) serve as the treated group, while graph-based problems requiring interactive platform manipulation serve as the comparison. This isolates AI-assisted offloading from general platform or curriculum effects.

Time-on-task decline

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

• High school: 31.3% cumulative decline
• Middle school: 9.0% cumulative decline
• Grade 5: no detectable change

The age gradient — absent for the youngest learners, strongest for high-schoolers and college students — is consistent with older students being more likely to access and use generative AI on their own.

Proctoring eliminates the effect

The divergence in study time vanishes entirely under proctoring for college students. This rules out general efficiency improvements, platform changes, cohort effects, or curriculum revisions as the explanation, and strongly implicates off-platform AI use as the driver. It is the cleanest evidence in the study that the decline reflects students substituting AI for their own cognitive work rather than learning faster.

Learning outcome impact

• Proctored retention items: a 25% cumulative decline in odds of correct response (logistic fixed-effects models on randomly assigned proctored retention items) — durable knowledge is measurably worse.
• Non-proctored assessment: a large opposite-signed increase — performance looks better when AI is available, but this is "impossible to attribute to anything other than AI assistance." This is the classic performance–learning gap: AI inflates immediate scores while eroding durable learning.

The "cognitive surrender" concept

The authors introduce 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, often metacognitively-managed strategy to a passive, unreflective dependency — echoing the knowledge base's distinction between adaptive and maladaptive offloading and the mechanisms documented under Reducing AI Misuse and AI Misuse and Learning Harm.

What this means for practice

• Learners. Notice the pattern this study names cognitive surrender and protect unassisted practice time: non-proctored performance rose while proctored retention items fell 25% in cumulative odds of a correct response.
• Instructors. Ground high-stakes Academic Integrity judgment in proctored, unassisted measures, because the post-ChatGPT time decline vanishes entirely under proctoring and the same estimator on non-proctored assessment yields a large opposite-signed increase that is impossible to attribute to anything other than AI assistance.
• Instructors. Redesign Assessment toward unassisted and process-based work, and build AI Literacy and Self-Regulated Learning training so students can recognize and resist passive dependency rather than only use the tools.
• Instructors. Do not treat self-report or non-proctored performance as evidence of learning: only proctored items exposed the 25% retention decline, so outcome measures must control for AI Accessibility.
• Administrators. Target AI Regulation in Education and AI policy by level and platform, since the population-scale, objective behavioral evidence shows cumulative declines of 31.3% in high school and 9.0% in middle school against no detectable change in Grade 5, and math mastery platforms like ALEKS are directly affected because text-based problems are the most AI-susceptible.

Limitations

• The learning-time analysis uses ALEKS learning data while the retention analysis uses the ALEKS PPL placement dataset — different populations under different conditions — so the individual-level causal chain cannot be established with the current data.
• AI use is never observed directly; all inferences come from behavioral signatures, validated by two falsification tests, and alternative behavioral explanations cannot be fully excluded.
• ALEKS PPL placement performance reflects prior learning, test-taking familiarity, and platform experience rather than purely retention of concepts practiced during ALEKS learning, so a gap remains between it and a laboratory-grade retention test.
• The retention analysis observes the mechanism only at the population level, even though the evidence is consistent with it.

Citation

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.