Block duplication in code rose 81% between 2023 and the present, climbing from 40.3 to 73.0 instances per million changed lines, according to a report published in June by GitClear. The analysis, titled the Maintainability Gap report, examined 623 million code changes across three and a half years. The findings suggest that as teams rapidly embraced AI development environments like Cursor and Claude Code, basic software quality practices deteriorated rather than improved.
The dataset reveals additional signs of declining code health. Moved code, which signals refactoring activity, dropped from 21% of altered lines in 2022 to just 3.8% in 2026. Heavy AI users increased their own prior output by 25%, falling well short of the 10x productivity gains frequently claimed by tool vendors. Those same heavy users out-produced colleagues without AI assistance by 4 to 10 times, but the report notes these teams were already outperforming their peers before AI tooling arrived. The data shows no clear connection between volume metrics like lines of code, pull requests, or feature counts and actual business value.
The report states that developers fundamentally changed their behavior after adopting AI tools. Before AI arrived, developers chose refactoring over copying and pasting code at roughly a two-to-one ratio. Now they're approximately five times more likely to duplicate code instead. The authors write that "the signals GitClear is getting from the data are a complete reversal: a return to the code-and-fix era of software development," referencing a bygone period when maintainability received little emphasis. The report finds that duplicated blocks of code drift apart over time, creating recurring bugs that developers must track down repeatedly.
The analysis explains that teams skip essential practices like test automation and refactoring because they feel pressure to demonstrate returns on AI tool investments. When organizations expect dramatic productivity jumps but see only modest 25% gains, developers cut corners to meet output targets. The report argues this creates an illusion of short-term progress while accumulating maintenance costs that become unbearable over time. Seemingly minor problems take far too long to diagnose and fix, with some issues simply becoming permanent because repair is economically unworkable. The accumulation of tightly coupled, incomprehensible code eventually reaches a point where the software stops delivering value. The report compares AI tools to a forklift truck rather than a racing car, noting that straight-line speed looks unimpressive but heavy lifts like replacing unmaintained libraries across a codebase represent untapped potential.
The report identifies high-performing teams as those that maintained rigorous technical discipline before AI tools arrived and continued those practices afterward. These teams monitor code health indicators and treat maintainability as fundamental rather than optional. The analysis recommends that organizations recognize technical practices like refactoring and test automation as core work, not side activities that can be deferred when deadlines loom. For software that people depend on and that needs to exist years into the future, professional delivery standards have moved beyond pick-and-mix approaches to a fixed set of non-negotiable practices. Organizations that haven't worked out how to measure value streams before purchasing AI tools will learn painful lessons when trying to track return on investment. The data suggests that winter is coming for AI tool budgets as companies balance their books and teams face harder usage limits where returns remain unclear.