Study: AI coding agents boost contributors but raise maintainability worries

A new preprint from researchers at the University of Notre Dame and other institutions provides the first large-scale empirical characterization of AI coding agents in open-source software. The team collected 13,360 AI conversation sessions (79,172 user messages) from 1,356 OSS repositories, linking chat data to development histories and supplementing with a developer survey. They found that AI use is heavier in smaller, less mature, and less collaborative repositories. After adopting AI, projects showed more active contributors and lower contributor concentration (p < .001), though communication remained highly concentrated. Code Writing was the dominant chat purpose, and nearly all AI sessions were followed by subsequent commits. Importantly, the study found no broad deterioration in code-quality signals or pull request merging rates, countering some fears about AI-generated code quality.

However, the survey uncovered a notable asymmetry: developers perceive others' AI-generated code as harder to maintain than their own (p = .029). While 68% of developers are willing to share their AI chat logs, concerns remain around appearing incompetent, increasing reviewer burden, and exposing ideas to competitors. The authors suggest these findings offer practical insights for designing and governing responsible vibe-coding practices in open-source development. The paper, currently available on arXiv, highlights both the opportunities and challenges of AI-assisted contribution, particularly around maintainability and trust in collaborative settings.