Cognitive Load in AI-Assisted Programming Environments: An Empirical Study

Abstract

The growing penetration of Artificial Intelligence (AI) into programming environments is reshaping software development, but its cognitive effects on developers remain poorly understood. In this paper, we empirically examine the influence of AI-assisted programming on developers’ cognitive load and task performance. We conducted a controlled within-subjects study with Baze University as a case-study, with more than 500 participants from mixed people of computer science students and software developers. Students solved standard programming problems (coding, debugging and comprehension) in both AI-supported and non-AI supported settings. Cognitive load (NASA Task Load Index) was assessed with objective performance data on task completion time, error rate and solution correctness. Results suggest that the AI-guided programming environments significantly lower the overall perceived cognitive load, especially mental demand and effort. Participants also performed better, solving tasks more quickly with better accuracy. However, qualitative and quantitative analyses show that the support by AI brings in additional verification and trust management work for experienced developers instead of replacing it, resulting in a shift rather than replacement of cognitive load. The most pronounced improvement in the cognitive demands were observed for students who received AI support and had: less difficulty mentally processing the text due to interference from topic (cognitive load here) and better understanding of what was being described (i.e., greater germane cognitive load). These results reveal a twofold cognitive role of AI-assisted programming tools as cognitive offloading devices and creators of new oversight demands. The study offers empirical contributions to human-centred software engineering research and highlights the need for AI-assisted programming environments with automation, transparency, and cognitive sustainability in balance.

DOI: https://doi.org/10.5281/zenodo.21541285