The Linux kernel has just put in writing something many maintainers had been muttering for months on the mailing lists: patches written by an artificial intelligence are no longer welcome in the project’s learning zone. Greg Kroah-Hartman, the kernel’s number two behind Linus Torvalds, published on Monday on the linux-staging list a new and blunt policy for the drivers/staging/ subsystem: from now on he will automatically reject any patch generated with a language model, with a single exception.
What exactly has changed
The announcement arrived on August 3 at 13:13 UTC and starts with a word that summarizes the problem: onslaught. Kroah-Hartman describes a recent wave of patches produced by language models aimed specifically at drivers/staging/, the directory where the drivers that still do not meet the main kernel’s quality standards live.
The decision is not a generic rejection of AI. Kroah-Hartman himself uses language models in his daily work with the kernel and has explained on several occasions that he finds them useful. What has changed is where they can be used: outside staging, AI-assisted work is still allowed, and Torvalds has recently reiterated that the Linux kernel “is not an anti-AI project”.
Why staging and not the rest of the kernel
The key lies in what that directory is for. As the maintainer himself explains, drivers/staging/ is “mainly a place for new developers to learn how to get involved in kernel development”. It is deliberately full of low hanging fruit: code cleanups, style tweaks, simple API changes. Low-risk tasks, perfect for someone to take their first steps without fear of breaking anything, to the point that loading any of those drivers activates the TAINT_CRAP flag in the kernel.
That is precisely why automatic tools are banned there. “We do not accept ‘fix all the style problems of this file’ type patches generated by tools, because that would destroy the very purpose of drivers/staging,” he writes. And he adds the nuance that explains everything: they could do that cleanup tomorrow if they really cared about the code of that directory, but they live with the flaws precisely because they are the training ground for newcomers. An AI that sweeps the directory in an afternoon does not fix a problem: it removes the exercise.
“It’s very obvious when someone sends an AI-generated patch”
The notice includes a warning unusual in tone for a technical list. Kroah-Hartman makes clear that not revealing the use of the tool is useless, because the style is recognizable at a glance, and that whoever is caught trying to “deceive” a maintainer deliberately has been warned. The goal, he insists, is that people learn, not that they pass a check.
The exception: real security, proven on real hardware
There is only one open door, and it is narrow. The maintainer acknowledges that current models have become very good at detecting possible security flaws in kernel code, but he accompanies the praise with a devastating figure: even with the best tools of this generation and the next, at least a third of the results they generate are directly wrong or harmful.
That is why whoever believes they have found a genuine flaw in staging with an AI can send it, but with two conditions: having tested it before on the driver’s real hardware, describing how they did it, and being willing to defend the submission and demonstrate that it fixes an error that a user can really suffer. The hardware filter is what should discard that third of false positives.
A debate that goes beyond the kernel
Kroah-Hartman’s policy is not an isolated case. Debian has been discussing for weeks a general resolution on the use of language models within the project, with five different proposals on the table ranging from allowing it with conditions to blocking it. Free software faces the same dilemma as newsrooms or universities: distinguishing the contribution that adds value from the one that only generates volume for someone else to review.
The maintainer closes his message with a metaphor that summarizes his stance better than any rule. drivers/staging/ is a gym, he says, a place you go to train and grow. A language model is becoming a reasonable tool for lifting weights, but that should only be done by those who have already trained enough to know what can be lifted and in what way.
