Linux 7.0 lands with two new documentation files that quietly close months of feuding on the Linux Kernel Mailing List over AI-assisted code. The rules live in coding-assistants.rst and generated-content.rst, and they are almost disappointingly sensible.

The core principle fits on a Post-it: AI is a tool. Use it if you want. The submitter signs for everything.

The actual rules

Every AI-assisted commit requires an Assisted-by: tag in the format AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2]. A patch that pulls in Claude and Coccinelle to refactor a driver carries:

Assisted-by: Claude:claude-3.5-sonnet coccinelle

Basic tools like gcc, make, and your editor stay out of the tag. Reasonable.

The critical part: AI agents must never add a Signed-off-by tag. Only a human can certify the Developer Certificate of Origin. The submitter takes full legal and technical responsibility for every line. There is no hiding behind "the model hallucinated."

The generated-content docs draw a clear scope line. Chatbot-written functions, Coccinelle-driven API migrations, AI-cleaned changelogs all need disclosure. Spelling fixes, identifier completion, clang-format runs do not. Maintainers get full discretion: accept, reject outright, demand extra testing, or (my favourite) suggest a better prompt.

The documentation puts it bluntly: "If you are unable to understand and defend everything you submit, do not submit the resulting changes."

Torvalds on the "AI slop" crowd

The sharpest commentary belongs to Torvalds himself. In January 2026, he wrote on LKML:

"There is zero point in talking about AI slop. That's just plain stupid. Why? Because the AI slop people aren't going to document their patches as such."

"The documentation is for good actors, and pretending anything else is pointless posturing."

He explicitly refused to make the kernel docs take any ideological stance: "I do not want any kernel development documentation to be some AI statement. We have enough people on both sides of the 'sky is falling' and 'it's going to revolutionize software engineering'." The man wants the patch to compile and be correct.

What compliance looks like in practice

Concrete walkthrough for a Claude-generated sorting function over a kernel linked list:

  1. Review the code for correctness, coding style, and GPL-2.0 compatibility
  2. Test with existing kernel infrastructure
  3. Add the human Signed-off-by
  4. Add Assisted-by: Claude:claude-3.5-sonnet
  5. Describe in the changelog what was AI-generated and how it was tested

Now contrast with the nightmare scenario Hackaday's commenters flagged: someone generates 50,000 lines of driver code, glances at it and submits without the Assisted-by: tag. Under the policy, maintainers can reject the entire series without review. As one commenter noted, "the average experienced engineer can only review about 500 lines of code in a change reliably." Generation speed and review capacity are now in completely different orders of magnitude.

Where Linux sits in the spectrum

NetBSD outright banned AI-generated code. Mesa added comprehension requirements after an AI slop incident. cURL shut down its bug bounty because AI-generated reports overwhelmed maintainers. Linux chose the pragmatic middle. And Neowin reported that Torvalds himself has been spotted vibe coding in personal projects, which tells you everything about where he actually stands.

The enforcement paradox

The policy works only on people who were already going to follow it. The top-rated HN comment on the 519-point discussion put it cleanly: "Basically the rules are that you can use AI, but you take full responsibility for your commits and code must satisfy the license. That's... refreshingly normal?"

Normal is the right word. Enforcement is the harder problem: MIT Media Lab research on AI-assisted writing cited in the same thread shows measurably lower cognitive ownership when humans delegate to LLMs, and the verification burden ends up on maintainers who already review more patches than any human should. Torvalds shipped the rules anyway, because pretending AI does not exist is even dumber.