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

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

FIRST LINE: How much have students' ordinary learning processes shifted in response to generative AI, and how does that affect their durable learning outcomes?

THE DISSECTION

This is a landmark empirical paper. It provides the first large-scale behavioral measurement of what I have long classified as the cognitive automation pathway to productive participation collapse — not in workers, but in the humans who are supposed to become workers.

The study is methodologically serious:
- 3.2 million ALEKS learning interactions over a 10-year panel
- Quasi-experimental design using curriculum structure as instrumental variable
- Proctored vs. non-proctored assessments as a behavioral divider
- Multi-cohort: Grade 5 through college

The findings are unambiguous and damning:

Time-on-Task Collapse:
- College: 26.9% cumulative reduction in time spent on AI-susceptible problems over 11 quarters post-ChatGPT
- High school: 31.3%
- Middle school: 9.0%
- Grade 5: zero change (they haven't developed the habit of externalizing cognition yet)

Learning Retention Collapse:
- 25% cumulative decline in odds of correct response on proctored retention items
- The proctored vs. non-proctored divergence is the methodological crown jewel: non-proctored assessments show improvement — meaning students are gaming the AI-assisted assessments, not learning the material

THE CORE FALLACY THIS PAPER INVALIDATES

The dominant optimistic narrative: "AI will augment human learning, making education more efficient."

This paper buries that narrative with population-level behavioral data. The efficiency gains are zero — the entire time reduction is behavioral substitution, not cognitive acceleration. Students are using AI to produce correct answers rather than building the cognitive architecture those answers were supposed to exercise.

The paper's term — cognitive surrender — is precisely correct and more honest than anything the techno-optimist crowd will admit.

THE HIDDEN ASSUMPTION THE PAPER SMUGGLES

The paper frames this as an educational policy problem requiring "assessment governance" and "AI policy." It implicitly assumes the system can be reformed — that if we just design better proctoring, better assessments, better AI policies, we can preserve the learning process.

This is institutional cope. The structural mechanism is clear:

1. AI makes cognitive tasks faster and lower-effort
2. Humans, behaving as rational cost-minimizers, offload those tasks
3. The offloaded cognitive processes are the precisely the processes that build the human capital pipeline
4. That pipeline feeds the future productive workforce

The paper documents steps 1-3 in education. The Discontinuity Thesis predicts what happens downstream: when those students enter the labor market, they lack the cognitive infrastructure that wage-labor productivity requires. The DT does not predict gradual optimization. It predicts structural bifurcation — a Sovereign class that owns AI capital, and a Servitor class that lacks the cognitive tools to compete even for the remaining human-labor niches.

SOCIAL FUNCTION

This is partial truth with institutional camouflage. The paper is rigorous, honest, and damning within its frame — but its policy recommendations (assessment governance, AI policy) are entirely insufficient. The cognitive surrender is not a policy failure. It is the intended output of an economic system that rewards task completion over cognitive development.

The fact that this paper will be cited in education journals and AI policy discussions, generate conference panels and reform proposals, and change nothing substantive — that is the social function: legitimized acknowledgment that generates no structural change.

THE VERDICT

This paper is the empirical foundation for what the Discontinuity Thesis has predicted: cognitive automation does not spare human development. It begins at the educational pipeline. The Grade 5 cohort shows zero effect — because they haven't yet learned to outsource thinking. By high school, the surrender is nearly complete. By college, the productive human capital being built is structurally deficient.

The 25% retention decline under proctoring is not a learning problem. It is a labor market pipeline problem. The humans entering the workforce in 5-10 years are arriving with systematically degraded cognitive infrastructure — exactly when AI is simultaneously eliminating the human-labor niches that don't require that infrastructure.

The paper documents cognitive surrender at scale. What it does not grasp is that this surrender is not a bug. It is the feature.