{"id":5589,"date":"2026-04-13T04:39:00","date_gmt":"2026-04-13T11:39:00","guid":{"rendered":"https:\/\/catbradley.io\/?p=5589"},"modified":"2026-04-13T04:39:00","modified_gmt":"2026-04-13T11:39:00","slug":"ai-code-gets-approved-in-the-linux-kernel-but-with-strings-attached","status":"publish","type":"post","link":"https:\/\/catbradley.io\/?p=5589","title":{"rendered":"AI Code Gets Approved in the Linux Kernel\u2026 But With Strings Attached"},"content":{"rendered":"

The Linux<\/a> kernel project has spent quite some time navigating the use of AI tools, and the response usually has been somewhere between “figure it out yourself<\/em>” and “we’ll get back to you<\/em>.”<\/p>\n

Late last year, at the 2025 Maintainers Summit, Sasha Levin<\/a> pushed for some documented consensus, and what came out of it<\/a> was human accountability for patches being non-negotiable, purely machine-generated submissions not being welcome, and tool use being disclosed.<\/p>\n

He promised to put something in writing without committing to enforce it, and that work has now shipped<\/a> with Linux 7.0<\/a>.<\/p>\n

What is it?<\/h2>\n
\n
\n
\n
\"\"<\/div>\n
\"\"<\/div>\n<\/div>\n<\/div>
\n

Linux’s <\/em><\/i>AI coding assistants<\/em><\/i><\/a> policy.<\/em><\/i><\/p>\n<\/figcaption><\/figure>\n

The new document is called AI Coding Assistants<\/a> and lives in the kernel’s process docs alongside the rest of the contribution guidelines. The short version is that AI-assisted contributions still need to comply with GPL-2.0-only<\/a>; AI agents cannot add Signed-off-by<\/code> tags; and patches that had AI help should carry an “Assisted-by<\/em>” tag.<\/p>\n

The Developer Certificate of Origin<\/a> (DCO) is an important aspect that exists, so there is a human accountable for every patch<\/strong>. AI assistance does not change that hard requirement.<\/p>\n

Basically, the human submitter reviews everything the AI produced, confirms it meets licensing requirements, and puts their own name on it with an appropriate mention that AI was used.<\/p>\n

The Assisted-by<\/em> tag format is Assisted-by: AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2]<\/code>; for scenarios where either single or multiple tools were used. The document gives Assisted-by: Claude:claude-3-opus coccinelle sparse<\/code> as an example.<\/p>\n

Back then, Linus was not even convinced a dedicated tag was necessary<\/strong> and suggested the changelog body would do the job. But now, the kernel community seems to have settled on the tag anyway.<\/p>\n

It’s already in use<\/h2>\n

We covered this earlier in the week<\/a>, but Greg Kroah-Hartman (GKH) seems to have had AI-assisted fuzzing running in his kernel tree for a while now, in a branch called “clanker<\/em>.” He started with the ksmbd and SMB code, found some potential issues, and submitted fixes with a note telling reviewers to verify everything independently before trusting any of it.<\/p>\n

That is just about the workflow the new policy was written around<\/strong>. AI surfaces issues, a human with decades of kernel experience decides what is real, writes the fix, and takes responsibility. GKH being the one doing it is not a surprise given he is the stable kernel maintainer and has probably dealt with more bad patches than the others.<\/p>\n

Other projects have gone in a different direction. Gentoo banned AI-generated contributions entirely<\/a> in 2024, with its council citing copyright risk, code quality, and ethical concerns.<\/p>\n

NetBSD’s commit guidelines<\/a> put LLM-generated code in the “tainted code<\/em>” category, requiring written approval from the core developers before any of it goes in.<\/p>\n

In contrast, Linux is not banning anything<\/strong>. Whether that turns out to be the sensible call or just a lenient one will depend on how seriously people actually take the “a human reviewed this<\/em>” part.<\/p>\n

\n

Suggested Read \ud83d\udcd6:<\/strong> Is a Clanker Being Used in Linux Development?<\/em><\/a><\/p>\n

<\/p>","protected":false},"excerpt":{"rendered":"

The Linux kernel project has spent quite some time navigating the use of AI tools, and the response usually has been somewhere between “figure it out yourself” and “we’ll get…<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-5589","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rss"],"_links":{"self":[{"href":"https:\/\/catbradley.io\/index.php?rest_route=\/wp\/v2\/posts\/5589","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/catbradley.io\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/catbradley.io\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/catbradley.io\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/catbradley.io\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5589"}],"version-history":[{"count":0,"href":"https:\/\/catbradley.io\/index.php?rest_route=\/wp\/v2\/posts\/5589\/revisions"}],"wp:attachment":[{"href":"https:\/\/catbradley.io\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5589"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/catbradley.io\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5589"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/catbradley.io\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5589"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}