One in ten Linux kernel patches says an AI helped

A line at the bottom of a patch

On September 14, Darrick Wong, who has contributed to the Linux kernel since 2006, sends in a fix for XFS, one of his filesystems. It repairs an in-memory copy that overflows its buffer and trashes whatever sits right behind it. At the bottom of the message, between his sign-off and the reviewer's, there's an Assisted-by line. It names a tongue-in-cheek tool and adds a comment: it was there to catch the obvious bugs.

That tag now appears on 1,877 of the 18,533 changes in version 7.3. Roughly one in ten. Since version 7.0, the kernel has asked developers to disclose AI use, patch by patch. And just as the counter becomes readable, the maintainers want to stop counting.

Where the number comes from

The idea of the tag goes back to a proposal by Sasha Levin, a developer at Nvidia, in July 2025, according to Next. The documentation around it landed in the kernel with version 7.0, released in April.

On October 12, Jonathan Corbet published LWN's account of the Maintainers Summit, an invitation-only day that brought about thirty kernel developers together in Prague on the 8th. It's where the figure comes from: 1,884 tagged changes through the seventh release candidate of 7.3, about 10%. Corbet adds right away that the real number of AI-assisted patches is surely much higher.

In the room, Dave Airlie, who maintains the graphics drivers, floated a very different order of magnitude: probably 90% of contributors now use LLMs, he said. That's an estimate tossed out in conversation and reported by LWN, not a measurement.

So the 10% is a floor. And it's a share of changes, not of lines of code. Nothing allows anyone to write that AI produced 10% of the kernel.

We recounted

Counting tags is a bit like counting umbrellas on a drizzly day: you learn less about the weather than about who's in the habit of carrying one. So we went back to the kernel's public history between version 7.2 and 7.3-rc7, published on October 11. We find the same 18,533 changes as LWN, and 1,877 tagged, seven fewer than its count.

This isn't the work of a handful of people: there are 433 different authors. The most prolific, an AMD engineer working on display drivers, signs 110 of them, or 6%. The top ten combined account for just 31%.

Nor is it a single tool. Claude shows up in 46% of tagged changes, OpenAI models in 16%, Gemini in 11%, Chinese models in 5%. And 19% just write "LLM" without naming anything.

The real shape of the data is elsewhere:

• They're regulars. Nearly nine tagged changes in ten (87.5%) come from authors who had already contributed to the kernel before 7.2. Several of the top ten are full maintainers who arrived between 2005 and 2007.
• They're mostly fixes. More than half (56%) repair a bug tied to the specific change that introduced it. For untagged changes, it's under a third (29%).

And the curve is steep: 31 tagged changes in 7.0, 294 in 7.1, 1,110 in 7.2, 1,877 in 7.3. From 0.2% to 10% in six months.

What the tag says, and what it doesn't

The rule is clear on one point: an AI never signs off. Only a human certifies where the code comes from, reviews what the machine produced and takes responsibility for it. The submission guide adds that forgetting the tag can slow a patch's acceptance. And the document on tool-generated content has the sentence that matters: you have to understand what you submit and be able to defend it. If you can't, you aren't submitting it.

What the tag doesn't tell you is the dose. Wong's LLM found the bug, and Wong fixed it. A whole file written by a machine carries the exact same line. It's the "may contain traces of nuts" label: it warns, it doesn't measure anything.

As of 7.3, it doesn't even name the tool anymore. In July, Christian Brauner got the rule cut down to a plain Assisted-by: LLM, arguing that naming models amounted to free advertising for proprietary vendors without adding useful information. And according to LWN, many contributors don't add the tag at all, while maintainers rarely check for it.

The counter being unplugged

In Prague, the room lined up behind a strong consensus: the tag has run its course, and the requirement will be dropped, LWN reports. Linus Torvalds says he's fine with that, and Brauner now considers it pointless.

The information would more likely go in the message that introduces a series of patches. As of 7.3-rc7, the documentation hasn't changed. Laurent Pinchart, for his part, objected on October 13 on the kernel mailing list, Next reports: as long as nothing replaces it, a rule at least sets expectations.

Torvalds moves the problem elsewhere. According to LWN, what worries him is patches from people who don't know the kernel: LLMs driven by contributors who don't know what they're doing. For a competent developer, he says, the tool is excellent, and it took him months to tune his own. He also thinks many people don't admit to using it out of embarrassment, and that a requirement could push some developers away from these tools.

Our recount points his way, from the other end: the people who declare are mostly regulars, the profile that doesn't worry him. The rest, by definition, show up in no counter.

That's the fault line already visible in the wave of AI contributions swamping maintainers. And while the share of code written by AI has never been measured, Linux was one of the few places where it was being declared.

The compiler nobody read

On October 15, The Register told the story of developer Theo Browne, who had an AI port the TypeScript compiler to Rust. A first attempt with OpenAI models burned through more than $400,000 in tokens at API rates, for about 84% compatibility. The second, with Claude Opus 5.5, cost around $24,000 and passes the conformance tests at close to 100%. Those are Browne's own figures.

The model ported the Go version published by Microsoft rather than the original code, and copied some of its bugs along the way. Browne only found out because someone else noticed. He says himself that he didn't inspect the code.

The two scenes don't differ over the tool. They differ over reading. The kernel is about to stop asking "with what?" It keeps the sentence that asks "do you understand it?" Browne answered the first question down to the dollar, and left the second open.