Exploring the Effect of Friction in AI Programming Tools on Cognitive Load
Authors
Jensen, Mike ; Kristensen, Rasmus Bisgaard ; Røn, Sebastian Malthe
Term
4. term
Education
Publication year
2025
Submitted on
2025-06-04
Pages
20
Abstract
This study examines whether adding friction to AI coding tools affects cognitive load, meaning the mental effort needed to understand and complete a task. Thirty-eight master's students in software engineering were assigned to two groups: a control group using standard GitHub Copilot (an AI code assistant) and a friction group using a modified version. The modified version slowed the acceptance of code suggestions, highlighted AI-generated code, and provided inline AI feedback on students' code. Participants completed programming tasks. We measured several aspects of cognitive load (intrinsic, extraneous, and germane; that is, effort from the task itself, from the tool/interface, and from learning), as well as confidence, frustration, and task progress. Data came from a questionnaire that also covered frustration and confidence, and from qualitative post-task interviews. Results showed no statistically significant differences between the friction and control groups on any measure: not in any of the three types of cognitive load, confidence, frustration, or progress indicators (such as tasks started or whether code compiled). Many participants overlooked or ignored the more subtle friction features. Opinions were mixed: some found friction disruptive, while others saw potential in educational contexts. Overall, the forms of friction we tested did not change cognition or behavior in short, time-pressured sessions. However, qualitative responses suggest possible benefits in other settings, including education.
Denne undersøgelse ser på, om det at indføre friktion i AI-værktøjer til programmering påvirker kognitiv belastning, forstået som den mentale indsats der kræves for at forstå og løse en opgave. 38 kandidatstuderende i software engineering blev fordelt i to grupper: en kontrolgruppe, der brugte standardudgaven af GitHub Copilot (en AI-kodeassistent), og en friktionsgruppe med en ændret udgave. Den ændrede udgave gjorde det langsommere at acceptere kodeforslag, fremhævede AI-genereret kode og gav inline AI-feedback på deres kode. Deltagerne gennemførte programmeringsopgaver. Vi målte flere sider af kognitiv belastning (intrinsisk, ekstrinsisk og germane; dvs. indsats fra selve opgaven, fra værktøjet/grænsefladen og fra læring), samt selvtillid, frustration og opgavefremdrift. Data kom fra et spørgeskema, der også dækkede frustration og selvtillid, og fra kvalitative interviews efter opgaverne. Resultaterne viste ingen statistisk signifikante forskelle mellem friktions- og kontrolgruppen på nogen mål: hverken i de tre typer kognitiv belastning, selvtillid, frustration eller indikatorer for fremdrift (fx opgaver startet eller om koden kompilerede). Mange overså eller ignorerede de mere subtile friktionselementer. Holdningerne var blandede: Nogle fandt friktion forstyrrende, mens andre så potentiale i undervisningssammenhænge. Samlet set tyder fundene på, at de former for friktion, vi afprøvede, ikke ændrer kognition eller adfærd i korte, tidspressede forløb. De kvalitative svar peger dog på mulige fordele i andre rammer, herunder uddannelse.
[This apstract has been rewritten with the help of AI based on the project's original abstract]