Linus Torvalds has strongly criticized attempts to restrict the use of AI tools in Linux kernel development, arguing that what matters is whether developers understand their code, not which tools they use to write it.
His comments came during a discussion on the Linux kernel mailing list, initiated by kernel developer Sasha Levin, over a proposal to remove the Assisted-by: LLM tag, currently used to identify patches developed with the help of AI coding assistants.
The tag was introduced earlier this year to give maintainers a way to identify AI-assisted contributions and review them accordingly. However, according to Levin’s revised proposal, the system has proved largely ineffective.
Of the 18,533 non-merge commits in Linux 7.3 through release candidate 7, about 10% included an Assisted-by tag, though AI tools are believed to be used much more widely.
Additionally, many contributors are unaware of the requirement, while others deliberately omit the tag. As a result, maintainers cannot reliably tell if AI was involved just by checking for it.
The proposed change would remove the requirement from the kernel’s contribution documentation and eliminate its handling in the checkpatch.pl script. Notably, Greg Kroah-Hartman, who previously supported keeping the tag, now supports its removal.
However, the discussion quickly became more heated when kernel developer Laurent Pinchart argued that clear expectations regarding AI use remain necessary, even if some contributors choose not to follow them.
As you might expect, Torvalds responded sharply in his usual style, rejecting the idea of restricting particular development tools.
“I don’t understand this whole idiotic concept of ‘don’t use AI for X‘,” he wrote.
He also sarcastically suggested developers should be forbidden from using compilers and required to write assembly code to prove they understand what their programs do.
“We’re not going to make stupid rules that say ‘you can’t use tools for X‘,” Torvalds stated.
In his view, developers should do their best and understand what they are doing, regardless of the tools, as competence cannot be judged simply by whether someone used AI during development.
Importantly, dropping the Assisted-by requirement would not mean abandoning transparency or accepting AI-generated code without scrutiny.
The existing documentation covering generated content would remain unchanged, and contributors would still be expected to take full responsibility for their submissions, including compliance with the Developer Certificate of Origin.
For now, the revised patch has received support from several prominent maintainers, but its removal of the disclosure requirement should still be distinguished from a completed, merged change.
In general, from the pure theoretical perspective, problem here that current (LLM) AI is not a Superintelligence, it is a Generative AI and is not just an Assistant it is a Tool for generating content (following by description/rules/tokens) and that content is completely untrusted by default even if “it just works”.
Do I mind if such stupid (in sense not really intelligent) tool is used by some clever persons like hardcore kernel maintainers – probably not so much because they can really see and understand generated code and properly adjust/review it (and can replicate it without AI that was used to just speed up routine).
But I do mind if kernel code (but any in general) can flooded by such kind of untrusted content that “just works” by the people that don’t really understand how exactly and why it works .
And I suppose “Assisted-by” supposedly should help reviewers to identify how code should be reviewed at first and if person delivered it was able to understand what was exactly was done and why.
Honestly the problem isn’t knowing why something works or not. That is what maintainers and the like are for to look over certain information and the like. When I first learned any kind of coding and programming it was looking over commented projects and the like and ai code is very good at commenting the code. This can help teach people. While I understand code if I read over it I cannot code nearly as fast or efficient as ai can just as a general rule.
Heck I wish I had had ai back when I first started to use Linux and was learning. When I first used desktop Linux skype was still pretty big and inline volume on USB powered headphones of any sort was not working. Funny enough I found out rather specifically after installing skype that it did.
I thrn went through all the different installed packages and libraries to find what made it work.