Paper Number
ICIS2026-1094
Paper Type
Completed
Abstract
AI coding assistants are increasingly embedded in software development, yet their effects on software quality may depend on where programming problems arise. Drawing on software modularity research, we distinguish between defects within modules, which involve localized implementation problems, and defects across modules, which involve dependencies across software components. Using proprietary longitudinal data from a large Chinese automobile manufacturer and a matched difference-in-differences design, we find that AI access is associated with a 32.1% reduction in defects within modules, while no statistically significant reduction is detected for defects across modules. Greater AI use intensity shows a similar pattern. Analyses for each interaction mode further show that Chat use is associated with fewer defects within modules, whereas Completion use is not significantly associated with either defect category. These findings identify software module boundaries and the distribution of information relevant to the task as important conditions shaping the quality effects of AI-assisted programming.
Recommended Citation
Chen, Yuyuan; Zhang, Yueyue; Ding, Yong; Zhang, Wei; and Zhang, Cheng, "The Boundaries of AI Coding Assistance: Within and Across Modules" (2026). ICIS 2026 Proceedings. 1.
https://aisel.aisnet.org/icis2026/hti/hti/1
The Boundaries of AI Coding Assistance: Within and Across Modules
AI coding assistants are increasingly embedded in software development, yet their effects on software quality may depend on where programming problems arise. Drawing on software modularity research, we distinguish between defects within modules, which involve localized implementation problems, and defects across modules, which involve dependencies across software components. Using proprietary longitudinal data from a large Chinese automobile manufacturer and a matched difference-in-differences design, we find that AI access is associated with a 32.1% reduction in defects within modules, while no statistically significant reduction is detected for defects across modules. Greater AI use intensity shows a similar pattern. Analyses for each interaction mode further show that Chat use is associated with fewer defects within modules, whereas Completion use is not significantly associated with either defect category. These findings identify software module boundaries and the distribution of information relevant to the task as important conditions shaping the quality effects of AI-assisted programming.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
13_Human_technology_interaction