Review and refactoring must scale with AI use.
What it is about
The authors examined around 302,600 verified AI-authored commits in 6,299 public repositories and tracked what happened to the issues they introduced. Result: 22.7% are still present in the latest version.
What caught our attention
Not the error rate itself - but that the issues stay. AI code appears to be revised less often than human code. Technical debt builds up quietly because the code looks plausible at merge time.
What we take from it
For us, a reason never to let agents work without a review gate and quality checks - and to plan refactoring as a fixed part of the work, not a leftover. How this looks in practice: From the artefact to the tool.
