GitClear never published the “40% of AI code is reverted” stat
The most-cited number about AI code quality does not appear in the report it is attributed to. I went and read the source. Here is what GitClear actually publishes, and the paper you should cite instead.
I almost put this in a slide. “GitClear found 40% of AI-written code is reverted or rewritten within two weeks.” You’ve seen it. It’s in blog posts, conference talks, LinkedIn threads, at least one VC memo I got forwarded.
I went looking for the source so I could cite it properly, and I couldn’t find one. So I stopped reading the blogs and read GitClear’s actual reports.
Here’s what they publish.
Their 2024 report, Coding on Copilot, covers roughly 153 million changed lines from January 2020 to December 2023. Their 2025 report covers 211 million changed lines through December 2024. Both define churn the same way: “the percentage of lines that are reverted or updated less than two weeks after being authored.”
The 2025 report has the actual table. Churn by year:
| Year | Churn (% of changed lines) |
|---|---|
| 2020 | 3.1% |
| 2021 | 3.3% (their stated pre-AI baseline) |
| 2022 | 3.3% |
| 2023 | 4.5% |
| 2024 projected | 7.1% |
| 2024 actual | 5.7% |
That’s the whole finding. Churn went from about 3% before AI tools to about 6% after. Roughly a doubling, off a small base. GitClear even walks back their own projection in the appendix: “Thankfully, short-term churn has not yet reached 7%.”
I searched both report PDFs for the string “40%”. Zero hits in each. Same for their 2025 and 2026 research pages. The chart everyone screenshots, on the 2024 report page, has no y-axis labels on it at all, which is probably part of how this got out of hand.
The only thing close to a 40 anywhere in their material is block duplication, which they report going from 40.3 in 2023 to 73.0 year-to-date — and that’s measured per million changed lines, not as a percentage of anything. I don’t know whether that’s where the number came from. I’m not going to pretend I do.
If you want a number for how much agent-written code gets reworked, there is a real recent paper. Popescu et al., Investigating Autonomous Agent Contributions in the Wild, arXiv 2604.00917, April 2026. 111,969 pull requests, five agents: Codex, Claude Code, Copilot, Jules, Devin. They measure at three days, one week, and three weeks. Not fourteen days.
Their result does look bad for agents at first. Human commits cluster between 0 and 0.4 on their churn ratio. Claude Code’s cluster between 0.8 and 1.
Then you read the next sentence:
“Mann-Whitney U tests show these differences are statistically significant (p<0.001), and Cliff’s δ values (-0.05 to -0.14) suggest the effect sizes are small but consistently negative.”
Small but consistent. That’s the honest version, and it’s a lot less quotable, which I think is the actual lesson here.
One caveat I want to be clear about, because it would be embarrassing to fix a bad citation with a sloppy one: the arXiv paper’s churn is churned lines over total lines in a commit. GitClear’s churn is the share of authored lines reverted inside two weeks. Those are different measurements. You can’t swap one number for the other, and I’m not offering the paper as a replacement figure. It’s a different, better-documented question.
I’m not writing this to dunk on anyone. The 40% was in my own draft for about a week. Nobody repeating it was being dishonest. It reads like a real number, it has a real research org attached, and checking it took me an afternoon that most people reasonably don’t have. If you’ve cited it somewhere, the fix is a two-minute edit.
Sources, so you don’t have to take my word for any of it
- GitClear 2025 report (churn table on p.7, definition in Appendix A0): GitClear-AI-Copilot-Code-Quality-2025.pdf
- GitClear 2024 report page (definition, 153M lines, “projected to double”): gitclear.com/coding_on_copilot…
- GitClear 2026 research (block duplication 40.3 → 73.0): gitclear.com/the_ai_code_quality_maintainability_gap
- Popescu et al. 2026: arxiv.org/abs/2604.00917