AI Is still making code worse: A new CMU study confirms (2025) (blog.robbowley.net)

A new study from Carnegie Mellon University, titled “Does AI-Assisted Coding Deliver? A Difference-in-Differences Study of Cursor’s Impact on Software Projects,” reveals that the integration of AI tools like Cursor has not improved code quality in open-source software projects. Analyzing 807 GitHub repositories using Cursor from January 2024 to August 2025, researchers found a brief spike in code generation during the initial adoption phase. However, this increase was not sustained, and the study reported a troubling 30% rise in static analysis warnings and a 40% increase in code complexity, indicating a decline in maintainability. This research is significant because it echoes earlier findings from GitClear, suggesting that despite advancements in AI-assisted coding tools, fundamental issues with code quality persist. The findings challenge the notion that poor AI-generated code is solely a result of user error, demonstrating that the tools themselves contribute to deteriorating code quality. As AI models are likely trained on increasingly complex and flawed codebases, the implications are concerning for future AI development, suggesting a potential feedback loop where worsening code quality reinforces itself in the AI learning process. The study underscores the ongoing responsibility of developers to ensure that code remains clear and maintainable amidst evolving technologies.