Skip to content

The Kernel Tried an AI Tag and Deleted It. What Replaced It Is Better

In October the Linux kernel removed its Assisted-by tag, after finding that about 10 percent of patches carried it while actual model use was far more widespread. Torvalds then stopped mentioning AI fixes in release notes. What is left is harder to fake than a label: a submitter who can defend the patch.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A matte black mechanical keyboard and a closed notebook on a dark desk under a single warm lamp, with an empty coffee cup beside them.
The label was removed because it stopped meaning anything. The obligation stayed.

Linus Torvalds got the attention on November 8, when he wrote in the Linux 7.4-rc2 release message: "And yes, much of this was driven by various AI fixes. I think I'll stop repeating that, and we can just take it for granted."

Fair enough, but the more interesting commit landed quietly on October 13 and almost nobody outside the kernel noticed it. Its subject line is "docs: remove Assisted-by references", and it deleted the kernel's AI disclosure tag.

The tag, and the numbers that killed it

The kernel added AI coding assistant documentation in December 2025, and that documentation asked contributors to add an Assisted-by trailer to patches developed with an LLM. In July 2026 the requirement was simplified: instead of naming the model, contributors would write Assisted-by: LLM, on the grounds that naming specific models is free advertising for the vendors.

Then came the maintainer summit, and this paragraph from the removal commit, which is the most honest piece of AI-policy writing I have read this year:

Roughly 10% of the 18,533 non-merge commits in 7.3 through -rc7 carry an Assisted-by tag, while actual LLM use is far more widespread. Some contributors do not know about the rule, others leave the tag off on purpose, and few maintainers ask about it. As a result, the tag does not provide the signal it was kept for: its absence says nothing about how a patch was made, and for developers who use these tools for almost everything, its presence says little more.

A disclosure mechanism that only appears on one patch in ten, in a codebase where model use is closer to universal, is not a signal. It is a coin flip with extra steps. The commit notes that the choice was between enforcing the rule and dropping it, that the room reached "a strong, though not unanimous, consensus" to drop it, and that Greg Kroah-Hartman, who had argued for keeping the tag in July, supported removing it by October. That last detail is worth sitting with. The person defending the rule changed his mind because the rule was not producing the thing he wanted.

What they kept instead

The kernel did not conclude that maintainers should stop caring how a patch was made. It concluded the changelog was the wrong place to say it. From the same commit: maintainers "do still want to know how a patch was created, but the view at the Summit was that this belongs in the cover letter, where it can describe how the tools were used".

That matches the guidance already sitting in generated-content.rst. It draws a line first: formatting passes, spellchecks, identifier completion and mechanical renames are out of scope. Anything more substantial is in scope, including the changelog itself if a model wrote it, and including a changelog translated from another language. Then it suggests, without mandating a form, what being transparent looks like: which tools you used, the input you gave them such as the Coccinelle script, the prompts if the code came from a short set of them, a summary of the prompts and the nature of the assistance for longer sessions, which parts of the content were affected, and how the submission was tested.

Two things follow from putting this in prose rather than a trailer. The first is that you cannot satisfy it mechanically. A tag is one line and a checkpatch rule, which is exactly why it decayed. A cover letter has to explain something. The second is that it gives maintainers something to act on, and the document is explicit about the menu: treat the contribution normally, reject it outright, review it with extra scrutiny, give it lower priority than human-generated work, ask how the model was trained, ask the submitter to explain the code in more detail, or "suggest a better prompt instead of suggesting specific code changes".

Under all of that sits the line that carries the entire arrangement: "You are expected to understand and to be able to defend everything you submit. If you are unable to do so, then do not submit the resulting changes."

Meanwhile, at the release level

Torvalds' November message is the other half of the same movement. In the rc2 notes he describes a large release, following the largest rc1 in kernel history, with a diffstat dominated by Coccinelle updates he had deferred from the merge window because he did not want to deal with them while travelling. Those scripted changes are mostly one and two liners spread across thousands of files.

Strip that noise out and the usual shape appears: a third drivers, mostly GPU, networking and sound; a third networking and BPF; a third "the rest", meaning filesystems, memory management, architecture. Then the datapoint that made him say something encouraging for once. Alexei Starovoitov told him that on the BPF side, fixes are now arriving faster than new reports, so the backlog there is shrinking instead of growing. Torvalds' response is careful: that "sounds lovely if true". Most other areas have not reached that point, and he signs off with "Don't hold your breath, but maybe cross your fingers we'll be getting there in the not entirely distant future."

Note what the BPF claim does not say. It compares fixes to reports. It does not say either was AI-assisted, and Torvalds does not add that. The honest summary is narrow: one subsystem with a tight contributor base may have crossed the point where incoming fixes keep up with incoming bug reports. That is worth watching. It is not a verdict on AI in open source.

What I would copy, and what I would not

The kernel's problem is not my problem, but the structure of the fix is portable, because every creative pipeline now has the same failure mode: a file exists, nobody remembers how it was made, and the person shipping it cannot explain the choices in it.

The kernel's list translates almost directly. Which model and version produced the image. The prompt, or a summary if the session ran long. Which regions were generated and which were retouched by hand. What was checked at export size, and with what. What you could not verify.

That last item is the one everyone omits and the one that builds trust. "Hands were not checked at 100 percent" is more useful to a collaborator than a green checkmark with no scope attached. The kernel document gives it a whole numbered step, and the reason is not politeness: unverified submissions burn reviewer time.

What I would not copy is the assumption that any of this survives on its own. The kernel's policy works because there is a person with commit rights who can simply refuse a patch. A creative team has no equivalent gate. If your asset pipeline accepts whatever lands in the folder, a provenance note is decoration. The reason the kernel's version has teeth is that the note is attached to a decision someone else makes.

One correction I had to make to my own reading: Signed-off-by does not certify that a human read the patch. It certifies the contributor's right to submit the work under the applicable license. The review obligation is a separate rule, which is why coding-assistants.rst puts it on the submitter as its own requirement, alongside the prohibition on agents adding that trailer themselves.

The kernel removed a label, and the thing that replaced it asks for more work, not less. That is the unusual part. Most disclosure regimes get weaker as they get easier to satisfy, and this one went the other way.


If you run a creative team, the cheap version is a three line note attached to every generated asset: what made it, what you changed, what you checked. It takes a minute. It saves the argument you would otherwise have six weeks later, when someone asks whether the version on the client's site is the one that was approved.

For the other side of agent delegation, the gym booking incident is a useful companion read: rules you never wrote are rules your agent will invent.

Sources

Share this guide
X LinkedIn