Linus Torvalds Draws the Line on AI Code Review Tools in Linux Kernel
Linus Torvalds declared Linux is not an anti-AI project and told critics they can fork or leave, drawing a line under LLM code review tools like Sashiko amid a shift from his previous skepticism toward AI-driven kernel tooling.
On July 14, Torvalds declared on the Linux kernel mailing list that Linux is not one of those anti-AI projects, instructing dissenting contributors they could either fork it or walk away, and closed a growing debate over automated tooling in the tree by drawing a line under how integration will be handled. The post arrived amid active discussion of Sashiko, an AI-powered agentic code review tool operating inside the kernel tree that has been generating both automated bug reports and algorithmic false positives. According to the mailing list thread, the system currently achieves roughly 53 percent True Positive rate for identifying actual defects, while its false positive output sits in the 20 percent range. Rather than demand architectural changes to dampen AI’s footprint in review workflows, Torvalds made it clear that dissenters unhappy with the pace of integration have two viable paths forward: maintain their own version of the codebase at their own expense, or simply step away from the mainstream tree.
The friction originates from a practical tension between traditional kernel review practices and automated evaluation systems. Sashiko runs continuously through the submission pipeline, scanning patches for logical errors, memory safety violations, and configuration conflicts before human maintainers even see them. A 53 percent True Positive rate means roughly half of its flagged issues are legitimate stability or security concerns that would otherwise require additional manual triage, while the remaining false positives create cleanup overhead that contributors must address. For years, kernel development operated on a highly selective, gatekeeper model where only experienced maintainers could safely merge high-risk patches without extensive review. Introducing an AI system that intercepts and evaluates submissions at scale requires maintaining those same safety standards while accepting a higher baseline of automated noise.
Torvalds has responded to the operational friction by treating it as a straightforward problem of evaluation rather than ideology. In stating he would “absolutely put my foot down as the top-level maintainer” in defense of AI tools, which he described as “clearly a useful one,” and emphasizing that kernel decisions will be based on “technical merit, not fear of new tools”, he framed the ongoing debate as a failure to separate performance metrics from platform anxiety. The kernel’s merging process already tolerates varying degrees of automated assistance, from compiler warnings to static analysis plugins, and Sashiko represents a continuation of that infrastructure shift rather than a departure from established norms. Contributors who flag false positives or question review quality are not asked to justify their resistance to AI itself, but to submit patched code or provide documented testing that demonstrates measurable degradation in kernel reliability compared to existing validation methods.
The current stance marks a distinct pivot in Torvalds’s public posture toward generative systems, particularly when applied to core open source development. In October 2024, he characterized 90 percent of AI marketing as hype and said he would deliberately ignore any claims that automated tools could meaningfully impact kernel architecture or contributor workflows. By mid-2026, that dismissal has been replaced by a pragmatic acceptance of AI-driven tooling, provided it demonstrably improves code quality without bypassing maintainers’ authority. The shift does not imply enthusiasm for the broader AI industry narrative; rather, it reflects an engineering calculus where any system producing even a modest true positive yield across millions of submitted patches earns its place in the pipeline so long as contributors can safely override or filter its output.
Whether Sashiko becomes a permanent fixture depends entirely on how contributors actually use it moving forward. A 53 percent True Positive rate is sufficient to justify continued development if review teams treat its output as preliminary rather than definitive, which aligns with how static analyzers and compiler suites have been integrated into the tree historically. Those who find the false positive volume manageable will continue merging patches, while those who consider any automated interference unacceptable can exercise the fork option Torvalds laid out. Kernel development has always operated by consensus among active maintainers, not by decree, but when top-level authority explicitly ties acceptance criteria to measurable technical outcomes instead of tool preference, the debate narrows to what actually works in production rather than how it was built.