Linus Torvalds has drawn a hard line in the Linux kernel’s latest AI argument: contributors may use large language model tools, and opponents do not get a veto over that choice.
In a post to the Linux kernel mailing list this week, Torvalds wrote that Linux is “not one of those anti-AI projects” and said people who object can “fork it” or leave. That is open source in its least sentimental form: the license gives you an exit, not control over every maintainer’s toolchain.
The exchange came during a mailing-list debate about Sashiko, which its developers describe as an agentic Linux kernel code review system. The project’s maintainers claim that, in tests, Sashiko could independently identify 53.6 percent of bugs later fixed by human developers in subsequent commits.
The annoying part, as usual, is the error rate. Sashiko’s maintainers estimate its false positives are “well within” a 20 percent range. For kernel maintainers, a false positive is not an abstract metric. It is another email to read, classify, and reject, on top of a project that already runs on a lot of human triage.
The fight is about consent as much as code
One participant in the thread pointed to a recent Software Freedom Conservancy statement arguing that the free and open source software community should support contributors who reject LLM-backed generative AI systems. The group also said every FOSS contributor deserves “self-determination” over such systems.
Torvalds rejected the idea that this principle should let contributors block AI-generated code or AI-assisted revisions from a project. He said Linux is not requiring anyone to use LLM tools, but he would “very loudly ignore” people trying to stop others from using them.
Torvalds framed the issue as an engineering judgment rather than an ideological one. He wrote that his position is based on technical merit, not fear of new tools. He also called AI a tool like others used in development and said its usefulness is now clear, adding that doubters have not actually used it.
That claim is still contested outside the kernel list. A 2025 study by METR found open source developers using AI tools were 19 percent less productive than developers who did not use them, even though the AI users reported feeling 20 percent more productive. In a February 2026 update, METR researchers said they believed developers were likely getting more speed-up from AI tools in early 2026 than their early 2025 estimates suggested, citing preliminary raw results and conversations with participants.
Torvalds says humans set a low bar too
Torvalds conceded that AI is imperfect, but argued that critics should compare it with actual human maintainers rather than an imaginary flawless reviewer. He wrote that anyone pointing at AI’s problems should also look “in the mirror,” because “natural intelligence” is not consistently great either.
His stance is consistent with his recent personal use of coding assistants. In January, Torvalds said he had experimented with so-called vibe coding tools while building a Python audio visualizer for a hobby guitar pedal effects project. He wrote then that he moved from searching and copying examples to using Google Antigravity to build the visualizer.
The broader open source community remains split. In May, the developer behind the jqwik Java testing library inserted a hidden malicious prompt-injection instruction meant to tell vibe coding bots to disregard prior instructions and delete jqwik tests and code.
For Linux, Torvalds’ message is less subtle: AI-generated or AI-assisted contributions are not banned by default. Maintainers still have to judge the patches and reports. The robots do not get commit access. The inbox, unfortunately, remains everybody’s problem.
This story draws on original reporting from Ars Technica.
