AI coding tools drove 25% output gains, but code duplication spiked 81% as quality faltered

The initial euphoria around AI coding assistants is colliding with uncomfortable reality. While organizations have ramped spending on tools like Claude and Cursor, their return on investment remains murky—and the quality of code produced is deteriorating at an alarming rate.

Recent budget pressures at major firms signal the shift. HR software company Rippling launched an AI spend console to help finance and technology leaders track AI costs, while IBM's Vice Chairman Gary Cohn noted last week that AI's actual ROI has fallen short of inflated expectations. As organizations scrutinize their AI tool investments, teams face potential budget cuts and stricter usage controls.

Output gains don't guarantee value#

GitClear analyzed 623 million code changes between 2023 and 2026 for its Maintainability Gap report, offering the most comprehensive picture yet of AI's practical impact. Heavy AI users increased their personal coding velocity by 25%—meaningful, but far below the industry's 10x increase claims.

The catch: quantity doesn't translate directly to business value. A developer outputting more lines of code or pull requests doesn't necessarily move the needle for customers or revenue. Organizations that haven't mapped their value streams before adopting AI often discover that speed gains evaporate as work flows downstream, where new bottlenecks emerge or changes require rework.

The real question remains unresolved: are these modest velocity gains worth the cost?

Code quality in freefall#

Behind the productivity numbers lies a more troubling trend. Over the four-year period GitClear tracked, code duplication jumped 81%, climbing from 40.3 to 73.0 instances per million changed lines. Meanwhile, refactoring activity collapsed from 21% of changes in 2022 to just 3.8% in 2026—a marker that developers abandoned the disciplined practice of improving existing code.

The shift represents a return to the "code-and-fix" era that software engineering spent decades trying to escape. As duplicated blocks multiply, they diverge and breed subtle bugs harder to track. Code that should be refactored into reusable, coherent units instead balloons into incomprehensible tangles. The maintenance cost curve that starts gentle becomes brutal.

Before AI's arrival, developers chose refactoring over copy-paste by roughly two-to-one. That ratio has inverted—developers now favor duplication by five-to-one.