Your AI coding spend bought 25% more output. Duplication rose 81%.

The latest argument over AI coding tools is getting less shiny. Rippling has already added a spend console for CFOs and CTOs, and IBM’s Gary Cohn said the ROI has not been nearly as high as many people think. Now The New Stack is pointing to a harder problem: even when AI lifts coding output, it may also make code worse to live with.

The clearest numbers come from GitClear’s Maintainability Gap report, published in June. It analyzes 623 million code changes from 2023 to 2026, which is a very large chunk of modern software work. On that data, heavy AI users improved their own prior velocity by 25%. That is real progress, but it is a long way from the 10x story that has been selling budgets and hype.

And the bigger catch is that output alone doesn’t prove value. The report says teams that outperformed their peers were often already ahead before AI tools arrived, and gains in coding speed can get swallowed by downstream work or shifted into new AI-driven tasks. If you count lines, pull requests, or features, the value picture still blurs fast.

On quality, the direction is uglier. Block duplication rose 81% over 2023, from 40.3 to 73.0 per million changed lines. Moved code, which is a refactoring signal, fell from 21% of changed lines in 2022 to 3.8% in 2026. Before AI, developers chose refactoring over copy-and-paste about two to one. Now they are roughly five times likelier to copy and paste.

That matters because software doesn’t just need to ship; it needs to stay understandable. The report argues that teams are drifting back toward code-and-fix habits, where short-term speed wins and maintenance gets kicked down the road. Eventually, that bill shows up as more rework, slower feature work, and bugs that are too expensive to untangle. AI may be helping teams move faster. It’s also helping them run up the debt ledger with impressive efficiency.