AI coding agents now contribute to software repositories at scale, but their documentation practices have remained largely uncharacterised. A new preprint on arXiv, titled "Documentation Drift in AI-Generated Pull Requests," aims to close that gap by examining a large corpus of AI-generated changes.
The study draws on the AIDev dataset, analysing 29,669 pull requests from 2,585 repositories. The abstract indicates the work measures documentation practices within these pull requests. The paper's title points to the core issue being examined: documentation drift, where documentation becomes out of sync with code changes.
The scale of the analysis—tens of thousands of pull requests across thousands of repositories—offers a broad view of how AI assistants handle documentation in real-world software development. The preprint was announced as a new arXiv submission, and its findings will be of interest to researchers studying AI-assisted programming and repository maintenance.