Linus Torvalds does not do subtle. When Laurent Pinchart argued on the kernel mailing list that the Linux kernel still needs a way to set expectations around AI-assisted patches, Torvalds told him to stop using compilers and write his submissions in assembler. He added that Pinchart should skip the text editor for his commit messages too.
It was sarcasm with a point. The October 13 reply landed in a thread about removing the Assisted-by tag, the label the kernel asked contributors to add when an AI tool helped write a patch. The Maintainers Summit had already agreed to drop the requirement. Torvalds explained the thinking behind that decision: the kernel cares whether you understand your code, not which tools you used to write it.
The short version
- The Maintainers Summit in Prague agreed on October 8 to deprecate the Assisted-by tag. Sasha Levin posted a patch the next day to remove the documentation that requires it.
- Only about 10 percent of patches in the Linux 7.3 cycle carried the tag, according to LWN, so it told reviewers little.
- Pinchart wanted a replacement disclosure mechanism. Torvalds rejected the framing and said he expects contributors to understand what they submit, whatever tools they use.
- Humans still sign off on patches and answer for every line. Maintainers still want to know how a patch was made, and the Summit suggested the cover letter as the place for that.
What the Linux kernel decided
The Maintainers Summit is an invitation-only meeting of roughly 30 kernel developers, held once a year. The 2026 edition took place in Prague on October 8. LWN’s report says large language models dominated the day, and by the end the room agreed the Assisted-by tag had outlived its purpose.
Sasha Levin, a kernel developer who works at Nvidia, posted a patch on October 9 that removes the Assisted-by references from the documentation. His note to the list said the Summit had reached consensus to deprecate the tag and that the text requiring it should go.
The thread passed 50 messages, with replies from Greg Kroah-Hartman, Kees Cook, Jonathan Corbet, Theodore Ts’o and others. New posts were still arriving on October 16. The Register reported the outcome on October 14.
Why the Assisted-by tag existed
The tag grew out of a year of argument over how the kernel should treat AI tools. The idea took shape in July 2025, and maintainers worked through the policy question at the 2025 Maintainers Summit in Tokyo. By early 2026 the guidance had reached the documentation tree as Documentation/process/coding-assistants.rst, which LWN says shipped with kernel 7.0.
The document set a few ground rules. AI agents must never add a Signed-off-by tag. The human submitter reviews all AI-generated code and adds their own sign-off to certify the Developer Certificate of Origin. Code must be compatible with GPL-2.0-only. Patches made with LLM help were supposed to carry an Assisted-by line.
Torvalds set the tone in January. When Lorenzo Stoakes pushed for language aimed at AI slop, Torvalds replied that documentation would never fix it, because the people sending slop do not label their patches. He wanted the guidance to treat AI as a tool and take no side in the wider fight.
Why maintainers gave up on the tag
Compliance, mostly. LWN counted the 7.3 development cycle through its seventh release candidate: 18,533 non-merge changesets, a record, and 1,884 of them (about 10 percent) carried an Assisted-by tag. LWN notes that the real share of AI-assisted work is surely much higher. At the Summit, DRM maintainer Dave Airlie guessed that roughly 90 percent of contributors use LLMs now.
Some contributors never read the rule. Others knew it and stayed quiet, partly out of fear of a backlash against LLM users. Few maintainers were asking submitters about AI use, so the tag carried no usable signal. Jonathan Corbet, who maintains the kernel documentation and wrote LWN’s Summit report, told the room the project should either enforce the requirement or drop it.
Torvalds thought many people were ashamed to admit they used the tools, and he worried that a mandatory tag might scare people away from them altogether. Christian Brauner called the tag pointless by now. Shuah Khan still found it useful, but the room’s consensus was that it had run its course. Thomas Gleixner half-joked that the kernel should add a Not-assisted-by tag instead, and Arnd Bergmann said he would find that a handy signal for deciding whom to spend time helping.
The tag also had critics before the Summit. In July, Red Hat’s Jeff Layton posted a patch to remove it, arguing that not everyone follows the policy, that the tag works as free advertising for proprietary LLM vendors, and that nobody had explained why the kernel needs the data in the first place.
The mood inside the project changed
Behind the paperwork sits a bigger shift. The Register notes that AI coding agents began producing code of acceptable quality for the kernel within the past year or so. In August, Greg Kroah-Hartman told an Open Source Summit audience in South Korea that “LLMs are good enough now,” and argued that they find bugs because code is made of patterns and LLMs match patterns. At the Summit he told maintainers to brace for a firehose of patches and bug reports for roughly the next 18 months, after which most of the bugs should be fixed.
Laurent Pinchart’s case for keeping disclosure
Not everyone agreed. Pinchart, a kernel developer and founder of Ideas on Board, said he was unhappy about removing the tag without a replacement mechanism for required disclosure. He feared submitters would push back whenever a maintainer asked about AI use.
His main argument was about expectations. A policy tells people what the kernel wants and does not want, he wrote, and that has value even if some contributors ignore it or cheat. The kernel has always had such people and has always rejected their work, so he did not accept that an unenforceable rule is a pointless one.
He raised a second worry at the Summit. LWN reports that Pinchart asked Torvalds to confirm the project is not trying to push out developers who choose not to use LLMs. Torvalds said it is not.
What Linus Torvalds wrote
Torvalds replied to Pinchart on October 13, and he did not soften anything. He opened by rejecting the framing outright, putting Pinchart’s motive down to personal hangups, not expectations. He called the whole notion of telling people not to use AI for a given job idiotic.
Then he built a parody rule. If tools are fair game for judging competence, compilers should be off limits too: anyone who cannot write kernel code in assembler clearly does not understand what the machine is doing. Text editors should go as well, because the backspace key lets you hide the mistakes you make while typing a commit message. He capped it with an all-caps swipe at how ridiculous rules like that would be.
The principle came after the sarcasm. The kernel will not write rules that forbid tools for a task. What it expects is that people do their best and know what their code does, and the tools have little to do with that.
I expect people to understand what they are doing.
He closed with a jab at the other side of the debate: the people he respects and considers smart tend to pick up new tools, not avoid them. The full message is archived on lore.kernel.org.
What Torvalds did not say
He did not say anything goes. In July he told the linux-media list, “Linux is not one of those anti-AI projects,” and added that people who object can fork the kernel or walk away. He also said nobody is forced to use AI tools.
At the Summit he pushed equally hard in the other direction. According to LWN, he argued that blaming LLMs points at the wrong target. The real problem, he said, is a growing number of patches from people who do not know the kernel or how to drive the tools. His advice was to give those submissions low priority, and he said the time has come to make contributing harder for newcomers. LWN counted nearly 800 first-time contributors in 7.3 alone.
Torvalds already uses AI on the kernel
The sarcasm lands harder given what he was doing that week. On October 11, a new branch was spotted in his kernel repository named noref-filemap-read-with-llm. It holds a patch series that lets the kernel read from the page cache without taking folio references, which removes a contention point when many threads read the same file. In the series description, Torvalds said the code, the test programs, the review and the first draft of the text were done with an LLM coding assistant, Claude Code, under his direction.
In his own benchmark, eight threads reading one file went from about 10 million reads per second to roughly 30 million, on his laptop with some security protections switched off. That is a narrow, contention-heavy workload and not a general speedup, and the series was not in the main tree in the latest reports.
A second piece of work appeared in the same repository: a tag named assoc-array-fixes. Torvalds explained that a reported bug sat in code too obscure for him to follow, so he used LLM tools to make sense of it. A user-space test harness confirmed the reported bug and found three more, all dating back to 2013.
The Summit conversation fills in the rest. LWN reports that Torvalds had played with an LLM on a separate project for months, but used one on the kernel for the first time only the week before the Summit. He took an old patch, had the tool revise it, and came away unwilling to work without one. He also said it took about six months to set the tool up so it stopped giving bad answers, and that he ends each session with a long cleanup pass.
That is a long way from 2024, when he described the AI industry as mostly marketing hype. At Open Source Summit Europe in Prague he told the audience he likes using AI, according to ZDNet’s report.
What changes for kernel contributors
- The tag goes away. Once the documentation patch lands, you will not need an
Assisted-byline. The patch was still under discussion on October 16, so treat the removal as agreed in principle and pending in the tree. - You still sign off. The guidance says AI agents must not add
Signed-off-bytags. You review the code, add your own sign-off and answer for it. Nothing in the reporting suggests that rule is changing. - Expect questions about how you built it. Maintainers still want to know how a patch was made. The Summit leaned toward notes in the cover letter instead of the changelog, and some maintainers would like to see which models were used and at what stage.
- Write your own changelog. LWN reports general agreement that even when an LLM writes the code, the human behind it should write the explanation.
- Low-effort patches wait. Torvalds and others said unknown contributors who send low-quality work should get low priority.
- Nobody has to use an LLM. Torvalds told Pinchart he is not trying to push out developers who do not.
How other projects handle AI code
The kernel’s answer sits between two camps. Gentoo voted to ban AI-generated contributions in April 2024, and NetBSD followed in May 2024 by declaring LLM-generated code presumed tainted and barred unless the core team approves. QEMU adopted a ban in 2025 and has since weighed a limited, disclosure-based exception for small fixes, tests and documentation. The kernel took a third route: allow the tools, make the human accountable and stop asking for a label.
The Register points out that the AI vendors do not agree on attribution either. Claude adds provenance messages by default, while Codex does not.
What to watch next
- The documentation patch. Whether Levin’s patch lands, and in which release, decides when the requirement formally ends.
- The cover-letter habit. The Summit suggested describing how a patch was made in the cover letter. Whether that becomes a norm or fades will show how much maintainers want the information.
- Torvalds’ page cache series. In the latest reports it sits on a branch and is not in the main tree.
- The unresolved worries. The Register notes that concerns about errors slipping through unnoticed, accountability and agents reproducing copyrighted code have not gone away.
Where this leaves the kernel
The Assisted-by tag failed for a plain reason: it depended on honesty, and few maintainers checked. What replaces it is accountability. A human signs, a human explains and maintainers decide how much time a patch deserves. Whether that holds up against a flood of LLM-driven submissions is the open question for the next 18 months.
Frequently asked questions
What is the Assisted-by tag in the Linux kernel?
It is a commit trailer the kernel’s AI guidance asked contributors to add when an LLM or similar tool helped create a patch. The requirement is being removed after the 2026 Maintainers Summit.
Is the Linux kernel banning AI-generated code?
No. The kernel accepts AI-assisted patches as long as a human reviews them, adds the sign-off and keeps the code compatible with GPL-2.0-only. Torvalds has said Linux is not an anti-AI project.
Do kernel contributors still have to disclose AI use?
Not through the Assisted-by tag once the documentation patch lands. Maintainers can still ask how a patch was made, and the Summit suggested putting that information in the cover letter.
Does Linus Torvalds use AI?
Yes. He built a page cache patch series with Claude Code, used LLM tools to untangle and test the assoc_array code, and told the Summit he no longer wants to work without an LLM.
Can an AI sign off on a kernel patch?
No. The guidance says AI agents must not add Signed-off-by tags. A human certifies the Developer Certificate of Origin and takes responsibility for the patch.
