- Author
- David Loker, CodeRabbit
- Published
- 17 December 2025
- Kind
- Company blog post summarizing a vendor study
- Original
- www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report
- Subject
- Coding with agents
What it is
David Loker posts on the CodeRabbit blog. CodeRabbit makes an AI code review tool. The post summarizes the company's report. The report compared 320 AI-co-authored and 150 human-only pull requests.
What it says
- CodeRabbit reports that AI-authored pull requests had about 1.7 times more issues overall.
- The post reports 10.83 issues per AI pull request, against 6.45 for human-only pull requests.
- It reports logic issues 75% more common, and readability issues over 3 times more.
- It reports security issues up to 2.74 times higher, and excessive I/O about 8 times.
- The post says no issue category was unique to AI; both make similar mistakes.
- The post recommends more context, policy-as-code, safety rails, and AI review tools.
What evidence it gives
The evidence is a count of issues found in 470 open-source pull requests. Counts are normalized to issues per 100 PRs. The post uses rate ratios to compare groups. Method details are in the full report, not this post.
Interests and limits
CodeRabbit sells a code review tool and the post ends with a pitch for it. The post says it labelled PRs as AI-authored by signals. It assumed the rest were human, and cannot guarantee that. The full report was not read; this reading covers the blog post only. Its explanations of why the patterns appear are stated without separate tests.