AI wrote 1 in 6 Linux kernel patches this month. Who reviews them?

On Sunday evening, @LundukeJournal posted "yet another record week for AI development of Linux." Last week, 1,634 patches submitted to the Linux kernel were written with AI, and so far in September AI-generated code makes up 17.25% of all kernel patches.

That's roughly one patch in six, in the software underneath most of the world's servers and every Android phone. When the number reached Hacker News overnight, one of the first replies was a skeptic asking "how would they even know?" Fair question.

How the count works

The count depends on a rule the kernel adopted in January. If you use an AI tool on a patch, you add an "Assisted-by:" line naming it, and Lunduke's figures come from those tags.

Back in May, when the weekly count first passed 400, the same account pointed out that the reporting is "based entirely on the honor system." That makes 17.25% a floor, since anyone who uses AI and skips the tag doesn't show up. It also counts submissions rather than merged code, and it can't tell you how much of each patch the AI actually wrote.

The trend is steep anyway. The first week of February had zero tagged patches, the weekly count passed 400 by the end of May, and last week it hit 1,634.

The rule that matters more than the number

I find the kernel's response more useful than the percentage. Its official guidance on AI coding assistants is short. AI tools must not add a Signed-off-by tag, the line that legally certifies where a contribution came from. Only a human can sign, and that human reviews the code, checks the licensing, and takes full responsibility for whatever it breaks.

So contributors can use any tool they like, as long as they label it and put their own name on the result.

Where it's straining

Writing code got a lot cheaper, and reviewing it didn't. In August, networking maintainer Jakub Kicinski said he and his co-maintainer were "completely overwhelmed," with a big share of incoming patches looking like low-priority AI fixes and cleanups (Dataconomy has the summary). They responded by pointing several frontier models at incoming patches as a first-pass reviewer, with support from Meta.

Greg Kroah-Hartman, who runs the kernel's stable releases, drew a harder line in one corner. The staging area is the part of the tree where newcomers learn on easy cleanups, and it now rejects LLM-generated patches unless they fix a real security problem. Even those have to be tested on actual hardware first. Kroah-Hartman uses AI himself. His objection is that a flood of machine-made style fixes crowds out the people staging exists to train.

My read

The biggest, most scrutinized volunteer code project in the world has settled on a few plain rules. Disclose AI use. Keep a named human accountable. Protect the reviewers' time, which sometimes means saying no to AI output that adds noise. The staging ban and the AI reviewers only arrived after maintainers were already overwhelmed, and I suspect most businesses will find their own rules the same way.

I expect the same squeeze in most offices. When a draft costs almost nothing, what runs short is someone qualified who has the time to check it.

What it means for a business

Your version of this is probably already happening. Someone on your team drafts quotes, customer emails, or job postings with ChatGPT or Claude, and the output went up. The time it takes to check that output stayed the same, and in most small businesses nobody has decided who does the checking.

The kernel's rules carry over almost word for word:

• Mark what AI drafted, even informally, so the reviewer knows where to look hard.
• Name one person who signs off on anything that reaches a customer, and make it their call.
• Budget review time the way you budget creation time. If AI doubles the drafts, someone needs hours to read them.
• Decide where AI isn't welcome. Kroah-Hartman did it for a training ground. You might do it for the work a new hire should learn by hand.

Working out where AI should draft and where a person should sign is part of what New Face Design's free process audit covers. We map who touches each piece of work before it reaches a customer and where the checking actually happens.