Linux Drops Its AI Label, but the Human Still Signs Off

For a while now, the Linux kernel has asked contributors to label any patch they wrote with help from an AI tool. That rule is on its way out. On October 14, documentation maintainer Jonathan Corbet applied Sasha Levin's patch "docs: remove Assisted-by references" to the docs-next branch, as heise reported. It still has to reach mainline in a future merge window, but the direction is set, and I think it's the right call.

A label that stopped telling anyone anything

The kernel's AI Coding Assistants guide asked for a line like Assisted-by: LLM coccinelle sparse in the commit message, and checkpatch knew how to check it. The point was a review signal. When Jeff Layton first proposed dropping the tag in July, Greg Kroah-Hartman pushed back: "all I want is the "signal" that someone is using a LLM so that I can review it as-such." That proposal failed.

By this month's Maintainers Summit in Prague, the numbers had caught up with the idea. Levin's revised patch says roughly 10% of the 18,533 non-merge commits in Linux 7.3 through -rc7 carry the tag, "while actual LLM use is far more widespread." A missing tag says nothing, and for people who use these tools for almost everything, a present one says little more. Levin describes "a strong, though not unanimous, consensus in the room to drop it." Miguel Ojeda checked the figure on the list and found tag use is actually rising, past 16% for work queued in linux-next.

Greg changed his mind, and said why

Some will call this a reversal, so let's be fair to Greg. He was never against AI as such. At KubeCon Europe in March, he told The Register that AI bug reports had gone from "AI slop" to real findings almost overnight. He had tried generating patches himself and found about two-thirds were right. "The tools are good," he said. "We can't ignore this stuff."

He did draw a hard line in one place. In August he announced he would automatically reject LLM-generated patches for drivers/staging, the corner of the tree that exists so newcomers can learn, with an exception for security fixes tested on real hardware. That is a judgment about what staging is for, not a ban on the tools.

On the tag itself, he did change his position, and he explained it plainly on October 13: "I've now changed my mind, it didn't work out at all and maintainers are really annoyed by it so let's just drop it and focus on the content of the patch itself (and push back on _ANYTHING_ you feel is wrong)." He then added his Reviewed-by to the new version. I'd much rather see a maintainer change course because the data says so than defend a rule that isn't working.

Not everyone is convinced

heise is right that the list is split. In the first thread, Laurent Pinchart argued that the missing tag gave maintainers "a low-effort way" to push back on suspicious series, and that the rules are about "setting our expectations, telling what we want and what we don't want." He suggested borrowing Mesa's rule that commit messages must be written by a human. Steven Rostedt, who missed the summit, said he would like a "Vibe-coded-with:" tag, because when he runs git blame he wants to know whether the author really wrote the code. Lorenzo Stoakes, who supports the patch, described reviewing AI slop on evenings and weekends to the point of burnout.

Linus Torvalds answered Laurent in his usual style, calling the "don't use AI for X" idea "idiotic." His bar: "I expect people to understand what they are doing. But that understanding has very little to do with what tools they use."

The human still signs

Here's the part that matters for accountability, and it didn't change. The guide still says AI agents must not add Signed-off-by tags, because only a human can certify the Developer Certificate of Origin. The submitter is still responsible for reviewing all AI-generated code and "Taking full responsibility for the contribution." Maintainers still want to know how a patch was made; that now goes in the cover letter, and the tool-generated content guidelines still ask for transparency.

Where I land

I use AI tools every day, and being against them outright makes no sense to me. A compiler is a tool, an editor is a tool, and so is an LLM. What worries me is vibe slop: people who can't read the code the model produced, never chose the tech stack, and never wrote a spec. Rostedt's worry is really about them. But a label never stopped that crowd, because they were the ones leaving it off.

What does stop them is what kernel maintainers already do: ask "how did you test this?", as Shuah Khan and Mark Brown suggested, and reject what the sender can't defend. Matthew Brost put it simply in the thread: don't send code you don't understand, and "never send code you haven't tested." Use the tools. Understand your work. Sign your name only when you mean it.