New Research On CoPilot And Code Quality

AI coding assistants are boosting short-term output while accelerating long-term code churn—an outcome that shows up in measurable changes to how open-source repositories evolve. The most striking finding is a shift in “churn” frequency: code lines are being changed far more often after being written, with churn that used to be low over roughly six months now appearing on an average cycle of about two weeks. That roughly doubles the turnover rate and compresses the time window in which developers must revisit and revise what they just produced.

The research behind these claims draws on a large dataset—about 21 million lines of changed code—tracking multiple dimensions of repository change rather than relying on productivity metrics alone. Across those dimensions, the pattern is consistent: developers are adding code faster, but they are getting worse at revising and refactoring it. The data points to more copy-pasted code, more duplicated blocks, and faster revision cycles, which together suggest that code is being committed and deployed before it has been properly validated or integrated into a maintainable structure.

A key “2024 milestone” is the balance tipping from refactoring to duplication. In earlier periods, “moved” code—refactoring or relocating code from one place to another—made up a larger share of change operations. By 2024, copy-pasted code surpasses moved code for the first time in the measured history. The implication is not that duplication never happens, but that the repo maintenance burden shifts toward co-maintaining multiple near-identical implementations of the same logic.

The transcript also sharpens what “copy-paste” means in this analysis. It includes cases where identical lines are duplicated within a single commit, not merely duplicated across a year of edits. That matters because it points to assistant-driven workflows where developers can accept multi-line suggestions quickly—sometimes with the AI effectively encouraging reuse by generating the same block again rather than calling an existing utility. The research further distinguishes “duplicated blocks” that are at least five lines long and already exist elsewhere in the repo; those large duplicated-block commits rose from about 0.5% in 2022 to 6.66% in 2024.

Churn-linked deletions are another warning sign. The most concerning deletions are “churn type” deletions: code that is newly authored and then deleted within about two weeks. The share of new lines changing within two weeks reached roughly 70% in the past year, while longer-horizon revisions (after one to two years) dropped to about a third of their earlier level. This aligns with defect-rate findings reported elsewhere: increased AI adoption correlates with higher defect rates, with one cited estimate suggesting a 25% increase in AI usage could raise defects by about 7%.

The discussion frames the problem as incentives and workflow, not just a single tool. If teams treat productivity as lines added or commits made, assistants that make it easy to “press Tab” will naturally increase the volume of code that later needs correction. The proposed remedy is cultural and technical: measure productivity in ways that reward reuse, consolidation, and maintainability; and for developers—especially juniors—differentiate by understanding the existing codebase well enough to reuse, refactor, and incrementally improve canonical implementations rather than recreating them in multiple places.