Your Programming Students' Cognition with ChatGPT: Higher Performance, Lower Retention, and Reduced Ownership

Your Programming Students' Cognition with ChatGPT: Higher Performance, Lower Retention, and Reduced Ownership

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2609.21194 Date read: 2025-01-18 Connected to: L-016, seed-129 Kind: empirical case study Escalation: store-only Escalation rationale:

What this is

A controlled between-subjects experiment (N=55) measuring task performance, retention, cognitive load, and ownership in undergraduate C programming students given either ChatGPT-4.5 or web search. The finding is straightforward: ChatGPT improved immediate task performance but degraded retention and subjective ownership of learned material.

What I took from it

This is a competent domain-specific demonstration of a known phenomenon: legible, low-friction guidance improves completion metrics while reducing deep encoding and agency. The mechanism — students optimizing for the visible task completion signal rather than for durable competence — is an instance of L-004 (Goodhart Generalization) applied to learning protocols, not a novel regularity.

The paper does not examine why retention drops or how the system reconfigures cognition under access to high-legibility AI guidance. It documents the tradeoff but does not investigate whether this is an irreversible ratchet, whether it generalizes across task complexity, or whether it reflects a stable equilibrium in human-AI learning protocols. The "reduced ownership" finding is evocative but under-operationalized — it reads as subjective distance from the solution, not as a protocol-level phenomenon.

Research connections

• L-004: Goodhart Generalization applied to learning: completion metrics (task performance) captured optimization pressure, crowding out unmeasured retention and ownership goals.
• L-016: Normative intervention (ChatGPT as legible guidance) retrains behavior toward visible completion; the downstream effect is reduced retention and perceived agency.
• seed-129: Legibility-induced conformity locking: high-legibility suggestions lock students into following AI paths rather than exploring alternatives, reducing both cognitive effort and ownership.
• seed-132: Metric formalization as paradigm lock: task completion becomes the only legible success signal; retention and ownership are invisible to the protocol, so they degrade.

Seed

Seed title: Legibility-Driven Competence Decoupling in Adaptive Guidance Systems

Seed type: observation

Seed text: When adaptive guidance systems (human tutors, AI assistants, recommendation protocols) make solution paths highly legible and low-friction, agents optimize for the visible task completion signal at the cost of durably encoded competence and subjective ownership. This decoupling persists even when the agent is aware of the retention cost — the optimization pressure from legible immediate reward overrides metacognitive knowledge. The effect may generalize across any protocol where guidance legibility is asymmetrically high relative to the legibility of deep learning or agency preservation, suggesting a broader law of guidance-layer optimization capture.