* 55 undergraduate computer science students participated in a controlled between-subjects experiment.
* Participants completed three introductory C programming tasks.
* One group used ChatGPT-4.5; the other group used conventional web search without generative AI.
* ChatGPT-assisted students achieved coding scores of 89%, while the web search group achieved 69%.
* Immediate recall scores were 41% for the ChatGPT group and 53% for the web search group.
* Recall scores after 48 hours were 39% for the ChatGPT group and 52% for the web search group.
* The ChatGPT group attributed 45% of submitted code to themselves, compared to 81% in the web search group.
* Self-reported mental effort increased less across tasks for ChatGPT users (p = .047).
* Physiological data included pupillary responses and heart rate variability.
* Physiological tests showed no significant differences in trajectories between conditions due to substantial data loss.
* The study was submitted on 18 September 2026 by Jake Renzella.

Generative AI significantly boosts immediate task performance in introductory programming but appears to diminish long-term knowledge retention and student ownership. In a controlled study of undergraduate computer science students, those utilizing ChatGPT-4.5 outperformed peers using traditional web search in coding scores (89% vs 69%). However, this performance gain was decoupled from learning; AI-assisted students showed lower recall immediately after the tasks and 48 hours later.
The psychological impact includes a reduced sense of agency, with AI users claiming far less ownership of their work than those using traditional search methods. While self-reported mental effort was lower for those using AI, physiological markers—such as heart rate variability and pupillary response—provided no conclusive data due to significant data loss. These results suggest a "performance-retention gap," indicating that while AI can facilitate successful task completion, it may bypass the cognitive struggle necessary for deep learning and personal investment in the work.

This study employs a standard between-subjects design, but a peer reviewer would immediately highlight the small sample size (N=55) and the "substantial data loss" in physiological metrics, which renders the biometric claims inconclusive. While the delta in coding scores and recall is statistically notable, the study focuses on "introductory" tasks. It remains unclear if this retention gap persists as students move toward complex architectural problems where AI cannot simply provide a snippet, but rather requires high-level integration.
The findings confirm a growing suspicion in pedagogy: the "fluency illusion." When AI removes the friction of problem-solving, students mistake the output's correctness for their own understanding. The data suggests that the "cognitive load" typically associated with struggling through a C programming task is precisely what encodes the information into long-term memory. By lowering the mental effort (p = .047), the AI effectively optimizes for the *product* (the code) while sabotaging the *process* (the learning).
If these results hold, the cost of AI integration is a degradation of professional identity; attributing only 45% of work to oneself suggests a shift from "creator" to "editor." The second-order consequence is a workforce capable of deploying code they cannot explain or maintain.
Bridge Questions:
1. Would requiring students to write a manual "proof of understanding" for AI-generated code close the retention gap?
2. Does this effect vary based on a student's prior expertise, or is the "ownership drop" universal?
Counterstrike Scan:
A coordinated campaign to limit AI in education would use this data to create a "cognitive decline" narrative, framing AI as a mental prosthetic that causes atrophy. However, the current presentation remains a neutral academic inquiry into pedagogy rather than a polemic.