Exploring the Effect of Friction in AI Programming Tools on Cognitive Load
Term
4. term
Education
Publication year
2025
Submitted on
2025-06-04
Pages
20
Abstract
We conducted an exploratory study to investigate whether introducing friction into AI programming tools can affect cognitive load. A total of 38 Software Engineering Master’s students were assigned to either a control group using standard GitHub Copilot or a friction group using a modified version with slow acceptance of code suggestions, highlighting of AI-generated code, and inline AI feedback of their code. Using programming tasks, a cognitive load questionnaire extended to measure frustration and confidence, and qualitative post-task interviews, we measured cognitive load, confidence, frustration, task progress, and qualitative reflections. Results showed no statistically significant differences between the friction and non-friction group across any metric, including intrinsic, extraneous, and germane cognitive load, confidence, frustration, or tasks started or compilation status. While participants often overlooked or ignored the more subtle friction elements, they expressed mixed views on the value of friction, some finding it disruptive, others seeing its potential in educational contexts. Our findings suggest that the types of friction we implemented do not significantly alter user cognition or behaviour in short, time-pressured settings, though qualitative responses point towards possible benefits in a different setting.
Documents