In which I explain how I will treat LLM generated or assisted patches for the Linux kernel drivers/staging/ subsystem going forward:
https://lore.kernel.org/all/2026080354-skater-urgent-31b2@gregkh/T/#u

(hint, not allowed, except for security bugs that you can prove actually fix something by testing the issue on the actual hardware the driver controls.)

@gregkh This post is how I learned that the kernel essentially has a playground for new contributors :D Nice!

@gregkh Is there anything to disincentivise people from trying until they catch you off-guard?

It's my impression that many of them don't care about being part of the community long-term, they just read on some blog that telling Claude to make kernel good is a great way to improve one's resumé and get a leg up in the rat race.

@gregkh good direction.

@CounterPillow As I said, it is very obvious when a newbie is using a LLM, so this is a warning that someone should not attempt to do that at all.

@gregkh let the hate flow through you it makes you stronger etc. ;)

@gregkh thank you. I hope it can tighten up further over time.

@gregkh could always just, you know, ban **all** LLM patches to the kernel, requiring a full human-rewrite + confirmed passing hardware test of any LLM output.

Kernel maintainers are among the heavy hitters of software engineering, responsible for the code at the core of the devices and services we all use every day. This is what one thinks about using a Trained #MOLE;

"Note, LLMs are very good at finding suspect 'security issues' in kernel code these days. But even with the best of the current and next generation tools, at least 1/3 of the results they generate are flat out wrong or harmful."

@gregkh, 2026

https://lore.kernel.org/all/2026080354-skater-urgent-31b2@gregkh/T/#u

#AI #LLM

I think the point is that people will use this to assist in throwing shit over the wall and making it your problem, or somebody else's problem once it's in. That said, bcachefs was dealt with quite successfully. However, I don't know if that works with small patches to things everyone depends on once you let it in.

@matthewcroughan @gregkh the point I was trying to make was more that if LLM patches to staging don't result in some negative consequences for the submitter such as automatic NAK replies on all their future submissions until trust is rebuilt, they'll just keep trying.

@gregkh That "at least 1/3 are wrong or harmful" line is the whole case against treating model output as a review pass. Finding candidates is cheap. Proving a fix on real hardware is the part that does not compress, and it is still the only part that counts.