Google Launches Sashiko, AI Code Review Tool for Linux Kernel
Key Takeaways
- ▸Sashiko detected 53% of bugs in recent Linux kernel patches that were 100% missed by human reviewers, demonstrating AI's value in code review augmentation
- ▸The tool has a false positive rate under 20% and is transparent about data handling, sending code to configured LLM providers like Gemini and Claude
- ▸Unlike AI-generated code submissions (controversial in open source), Sashiko takes a non-controversial approach by assisting human reviewers rather than replacing them
Summary
Google engineer Roman Gushchin has announced Sashiko, a new AI-powered code review system written in Rust that screens patches submitted to the Linux kernel for bugs and potential issues. The tool was developed internally at Google and works by ingesting patches from mailing lists, analyzing them with LLMs (primarily Gemini 3.1 Pro), and providing automated feedback to maintainers and developers.
In benchmarks, Sashiko detected 53% of bugs in a set of 1,000 unfiltered recent upstream issues—and notably, 100% of those bugs were missed by human reviewers during their initial review. While the 53% figure might seem modest in isolation, the fact that it catches issues human reviewers overlooked demonstrates genuine value. The tool's false positive rate is estimated at under 20%, with most borderline cases falling into a "gray zone." The authors acknowledge privacy and data-sharing considerations, noting that Sashiko sends code to configured LLM providers and works with Gemini Pro 3.1, Claude, and other models.
Belonging to the Linux Foundation, Sashiko represents a pragmatic middle ground in the heated debate over AI's role in open source. Rather than generating code submissions (controversial among maintainers), it augments the human review process, potentially easing the burden on kernel maintainers overwhelmed by review requests. Google is currently funding the tool's operation for the Linux Kernel Mailing List.
- The Linux Foundation–owned tool has been validated internally at Google and shows promise for easing maintainer burden across the open source ecosystem
Editorial Opinion
Sashiko exemplifies the most constructive path for AI in open source: augmenting expert human judgment rather than replacing it. Catching bugs that 100% of human reviewers missed is remarkable, even at 53% coverage, and shifts the conversation from "AI is stealing our jobs" to "AI is making us better reviewers." By staying transparent about data sharing and false positives, and by securing funding from Google, Sashiko avoids the antagonism that plagued earlier AI-in-open-source efforts. If it delivers on its promise, this model—AI-as-reviewer-aid—could become a template for responsible AI integration across the Linux ecosystem.
