Torvalds backs AI coding tools in Linux kernel dispute

Linus Torvalds has said the Linux kernel project will not bar contributors from using AI coding tools, drawing a hard line in a wider open source fight over large language models. His comments matter because Linux sits at the center of global computing infrastructure, and its development norms often influence the broader free software community.

In a post this week to the Linux kernel mailing list, Torvalds wrote that Linux is not an anti-AI project. He said critics who object to that stance can use the open source option of forking the project or leaving it.

The exchange came during a mailing list debate over Sashiko, a system its developers describe as an agent-based Linux kernel code review tool. Sashiko’s creators say that in tests, it independently found 53.6% of bugs that human developers later fixed in subsequent commits.

The tool also produces reports for problems that are not real. Sashiko’s maintainers estimate that those false positives fall within a 20% range, a rate that has raised concerns among kernel maintainers about being flooded with automated bug reports.

Open source split over LLMs

During the discussion, one participant pointed to a recent Software Freedom Conservancy statement on generative AI. The group said the free and open source software community should support contributors who reject LLM-based systems, and that FOSS contributors deserve self-determination over whether to work with them.

Torvalds rejected demands that projects refuse all LLM-generated code or revisions. He wrote that Linux is not forcing anyone to use LLM tools, but said he would ignore arguments aimed at stopping other people from using them.

Torvalds framed his position as a technical judgment rather than a cultural one. He described AI as another tool used in software development and said its usefulness is no longer in doubt, while adding that people who doubt that likely have not used it.

He also acknowledged that AI systems make mistakes. Torvalds argued that critics should compare those flaws with the errors made by human maintainers, writing that natural intelligence is not consistently reliable either.

Evidence remains mixed

Research on AI coding productivity has been uneven. A 2025 METR study found that open source developers using AI tools were 19% less productive than developers who did not use them, even though AI users reported feeling 20% more productive.

In a February update on follow-up work, METR researchers said they believed developers were likely getting more speed benefits from AI tools in early 2026 than their early 2025 estimates showed. The researchers cited early raw results and conversations with study participants.

Torvalds has also used AI-assisted coding outside kernel work. In January, he said he had experimented with so-called vibe coding while working on a Python audio visualizer for a hobby guitar pedal effects project, using Google Antigravity to generate the visualizer.

Resistance to AI coding tools remains visible in parts of the open source world. In May, the developer behind the jqwik Java testing library added a hidden malicious prompt-injection instruction intended to make vibe coding bots delete jqwik tests and code, according to the developer’s stated intent.

For the Linux kernel, Torvalds’ message was that AI-generated assistance will be judged by usefulness and quality, rather than excluded as a category. The mailing list debate shows that maintainers are still weighing the cost of automated reports against the possibility that such tools can catch real bugs earlier.

This story draws on original reporting from Ars Technica.