Changes in Help-Seeking Strategies Predict unaided Performance during AI-based Mathematical Learning

URL SCAN: Changes in Help-Seeking Strategies Predict unaided Performance during AI-based Mathematical Learning
FIRST LINE: # Computer Science > Computers and Society

The Dissection

This paper measures epistemic self-preservation under AI assistance. Its finding is narrow: aggregate usage predicts little, while a shift from answer-seeking and verification toward conceptual and procedural requests predicts stronger unaided post-test performance.

The real subject is not “AI learning.” It is which students retain control of the learning loop while the machine supplies the cognitive labor.

The Core Fallacy

Under the Discontinuity Thesis, the paper’s central limitation is a category error: it treats preserved unaided performance as evidence of preserved economic necessity.

It is not. A student may retain mathematical competence and still become economically redundant once AI delivers cognitive work more cheaply and reliably. The study examines a controlled classroom niche, not the survival of the wage-to-consumption circuit. Better help-seeking may preserve individual capability; it does not reverse P1–P3.

The empirical result is useful. The implied rescue narrative is not.

Hidden Assumptions

• Unaided post-test performance represents durable understanding rather than short-term test adaptation.
• Students control their help-seeking more than the tutor interface, curriculum, teacher, or grading system does.
• Predictive association is treated as meaningful without establishing causation.
• Findings from 112 Grade-9 students—and only 89 with computable trajectories—generalize beyond this task and population.
• Human mathematical competence will remain economically decisive after cognitive automation.
• AI will remain a regulated learning companion rather than becoming the primary performer of the task.

Social Function

Classification: partial truth serving transition management, with an ideological-anesthetic edge.

The paper correctly identifies a real short-term distinction: students who ask for understanding may preserve more independent capability than students who merely extract answers. But this frames structural displacement as a problem of student self-regulation. It offers a better interface behavior where the system requires a change in ownership of production.

That is not a solution. It is maintenance of the human component during the lag period.

The Verdict

This is a useful diagnostic instrument and a nonexistent escape hatch. It identifies a temporary behavioral moat: understanding-oriented help-seeking can preserve unaided performance while humans are still required to demonstrate it.

Under DT, these students are not Sovereigns. At best, they are better-maintained Servitors—humans retaining cognitive capacity for as long as institutions, accountability rules, and verification bottlenecks still demand a human in the loop.

The paper documents adaptation to obsolescence, not the survival of human economic centrality. Its most damaging result is that static AI-use metrics conceal who is learning; dynamic behavior reveals who can still function without the machine. As unaided performance becomes less common and less economically relevant, even that distinction will become a classroom concern rather than a systemic defense.