Does AI-Assisted Coding Deliver? A Study of Cursor's Impact on Software Projects (arxiv.org)

Researchers tested whether using an LLM-powered coding assistant (Cursor) actually boosts real-world software development by applying a causal difference-in-differences design to GitHub projects: Cursor adopters were compared with a matched control group of similar projects that didn’t use the tool. They find a clear, large but short-lived jump in project-level development velocity immediately after adoption. Crucially, Cursor adoption also produced a significant and persistent rise in static-analysis warnings and code complexity—quality degradations that didn’t disappear after the velocity bump faded. Using panel generalized method-of-moments (GMM) estimation, the authors show that the increase in warnings and complexity is a major mediator of longer-term slowdowns in velocity, suggesting a trade-off where short-term productivity gains are offset by enduring maintenance costs. Methodologically robust (matched controls, diff‑in‑diff, panel GMM), the study provides causal evidence that popular LLM agents can change development dynamics but may harm code quality and long-term productivity. Implications span practitioners (beware tool-driven technical debt), tool designers (prioritize quality-preserving features like better linting, test generation, and safe-suggestion filters), and researchers (need for longitudinal, production-scale evaluations of developer-facing LLMs).