Short: Greg Kroa-Hartman stopped accepting in the Linux kernel staging branch patches generated by neural networks — due to a flood of low-quality auto-submits. We analyze what happened and what rules this imposes on everyone writing AI-assisted code.
Greg Kroa-Hartman — the second person in the Linux development hierarchy after Torvalds — effectively closed the door for patches generated by neural networks to the staging subsystem of the kernel. The news echoes the recent discussion of “meat proxies” on the forum, only now the topic is not about a working chat, but about one of the largest open-source projects on the planet. We’ll unpack what happened exactly and what lessons to take for anyone writing code with AI.
What happened
Staging is a special “antechamber” of the kernel: raw drivers and modules go here that are not yet ready for the main tree. Historically this is a place for newcomers: a place where a beginner can contribute their first patch and learn the process under the supervision of maintainers.
It is precisely through this entry point that the wave of AI-generated code hit. According to Kroa-Hartman, staging was flooded with a stream of patches from large language models — formally written but of low quality, and there was no one to clear them up. The response: patches generated by LLMs in staging are no longer accepted. An exception is made for truly valid security fixes — they will be considered.
A telling detail: Greg himself is not a Luddite and successfully uses AI in kernel work. So this is not a “ban on neural networks” in principle, but a protection against a concrete problem — a flood of thoughtless auto-submits, each of which must be checked by a real person.
Why this matters (and not only for kernel folks)
This is the same economy as in the meat proxy story, just on the scale of thousands of contributors. Generating a patch now takes seconds. But verifying it remains expensive: a maintainer must read the diff, understand the subsystem context, assess correctness and safety. When generation is free and review is not, any open submission is instantly flooded, and the cheapest part of the work (click “generate and submit”) pushes the most expensive (figure out if this is correct) onto the person on the other side.
An irony here is that the entry point for beginners is precisely the one that suffers. The purpose of staging is to grow a developer who understands what they commit. The AI pipeline breaks this: a person submits code they do not understand, learning does not happen, and the maintainer gets noise instead of a learner.
A sober view
It’s important not to fall into “AI is evil.” Neither Greg nor common sense say that. What is worth documenting:
- The problem is not the tool, but irresponsible use. Kroa-Hartman uses LLMs — with understanding and review. The ban targets auto-submits, not the technology.
- This is not a universal kernel rule but a decision for a specific subsystem and a specific flood of junk. Other maintainers set their own policies.
- Security fixes are an exception, which makes sense: the cost of letting an error through is higher than the cost of verification.
Takeaways for you if you write AI-assisted code
The rules are simple and applicable to any pull request, whether in the kernel or your pet project on the forum:
- Don’t submit what you don’t understand. Your ability to explain every line of your PR is the minimum threshold. If you can’t, don’t submit yet.
- You, not the model, are the patch author. Responsibility for correctness lies with the person behind the commit. “An AI wrote this” is not an excuse for the reviewer.
- Respect the maintainer’s time. One well-thought, tested PR is more valuable than ten generated “just in case.” Mass auto-submits are a direct path to a ban.
- Test on a real build. A patch that looks correct but isn’t built and tested is work shifted to others.
In essence, the Linux kernel has just articulated to the world what every AI-enabled team discovers: the model is a booster for a competent developer, not a substitute for responsibility. As soon as a human drops out of the chain as a thinking filter, the AI pipeline becomes a noise generator, and communities start defending themselves.
Sources
- Phoronix analysis: Linux staging rejects LLM patches
- Our previous discussion on the topic: “Don’t be a meat proxy”
Where do you think the boundary for acceptable AI in open source lies: labeling generated code, limiting auto-submits, or relying on contributors’ common sense? And would you accept LLM patches in your project?