How Should the Linux Kernel Handle AI-Generated Contributions?

Linux kernel maintainers "are grappling with how to integrate AI-generated contributions without compromising the project's integrity," reports WebProNews: The latest push comes from a proposal by Sasha Levin, a prominent kernel developer at NVIDIA, who has outlined guidelines for tool-generated submissions. Posted to the kernel mailing list, these guidelines aim to standardize how AI-assisted patches are handled. According to Phoronix, the v3 iteration of the proposal [posted by Intel engineer Dave Hansen] emphasizes transparency and accountability, requiring developers to disclose AI involvement in their contributions. This move reflects broader industry concerns about the quality and copyright implications of machine-generated code.

Linus Torvalds, the creator of Linux, has weighed in on the debate, advocating for treating AI tools no differently than traditional coding aids. As reported by heise online, Torvalds sees no need for special copyright treatment for AI contributions, stating that they should be viewed as extensions of the developer's work. This perspective aligns with the kernel's pragmatic approach to innovation. The proposal, initially put forward by Levin in July 2025, includes a 'Co-developed-by' tag for AI-assisted patches, ensuring credit and traceability. OSTechNix details how tools like GitHub Copilot and Claude are specifically addressed, with configurations to guide their use in kernel development... ZDNET warns that without official policy, AI could 'creep' into the kernel and cause chaos...

The New Stack provides insight into how AI is already assisting kernel maintainers with mundane tasks. According to The New Stack, large language models (LLMs) are being used like 'novice interns' for drudgery work, freeing up experienced developers for complex problems... The Linux kernel's approach could set precedents for other open-source projects. With AI integration accelerating, projects like those in the Linux Foundation are watching closely... Recent kernel releases, such as 6.17.7, include performance improvements that indirectly support AI applications, as noted in Linux Compatible.

Burn it with fire

[Anonymous Coward]

And then bury it under 6 feet of reinforced, prestressed concrete.

It's a Tool

[dbialac]

My experience with AI has been that it's a tool that gives you a starting point, not a fully worked out solution to your problem. This cuts across multiple fields I've used it for. You need to take the approach verifying everything. And if you don't understand the field well, don't trust until you do.

Re:It's a Tool

[AmiMoJo]

Exactly. Treat it like any other submission. If it is well written, clear, works as intended, passes all the tests, then it's good. Who or what wrote it isn't really relevant.

Re:

[postbigbang]

It's important to also view contributions against a cumulative set of bad intention, woven as iterative contributions towards a nefarious goal.

Nonetheless, some may be benign and helpful. Others may, as an aggregation, be more onerous and "bend a branch" in ways unintended by the goals of the project. This could be said of human contributions, too.

Re:

[serviscope_minor]

Sure, but there's an additional problem which is that AI is very good at generating convincing looking PRs that turn out to be junk. The result is it can be quite a lot of work to figure out how junky the PR is. It kind of falls into the category of "and this is why we can't have nice things". There's nothing wrong in principle with submitting a PR with AI assistance, if the PR is sound. But unfortunately people looking to get their name on the kernel, for props or just frist psot will flood the mailing lis

Re:

[allo]

You know how Torvalds sounds when a PR is junk? He catches so much human junk, AI junk will be no problem for him.

Re:It's a Tool

[StormReaver]

Treat it like any other submission.

You can't treat it any other submission, for a variety of reasons. Some important ones are:

1) AI did not create the code. It lifted it from somewhere else, and it will be contributed without proper attribution.
2) There is no way to know if the lifted code has a compatible license, since there is no way to know where it came from.
3) If (or when) the rightful owner is identified, the lawsuits will fly.

Nuke it from orbit, then destroy the galaxy from which it originated. It's the only way to be sure.

Re:It's a Tool

[AmiMoJo]

All those things could be true about human generated code too.

Re:

[StormReaver]

All those things could be true about human generated code too.

Humans are answerable for their misdeeds. LLMs are not.

Re:

[AmiMoJo]

It would be the human who submitted the AI code who is on the hook.

Re:

[Ksevio]

There's no precedent for AI generated code retaining the license of code from the training set. There are only so many ways to write some of this code, it's bound to show up even if someone independently writes the same code.

The amount of code in a typical PR for the kernel would likely be more unique for the context. It wouldn't have a "rightful owner", it's just code how it's written.

Re:

[allo]

1) That's not how it works
2) So it is irrelevant
3) And that will not happen.

Let's say you would have the risk of overfitting (not that likely for code for several reasons), than it would be a risk for "Write me a quicksort", but it wouldn't not be a risk for "Write me a device driver for this spec" as such a driver didn't exist before. But when you look at a typical programmer workflow, that's not how AI is used. It is more like "Add asserts to the invariants of the selected loop and suggest if the order of

Re:

[dbialac]

made by aliens from the planet Zod

I dunno. I've met some aliens from the planet Zod. They can only travel through space because they were given the technology by another species, the Klorats. They're so inept that they once crashed a space ship into Roswell, NM when they were trying to meet with US military personnel at Area 51. As such, I wouldn't trust a line of code if they wrote it.

Can the dev explain how it works?

[DjangoShagnasty]

If not: bin it.

On the internet...

[greytree]

On the internet, no-one knows you're an AI.

https://i1.kym-cdn.com/photos/images/facebook/000/427/569/bfa.jpg

AI code = Public Domain

[Kisai]

That is how it's been, Those AI tools were trained on open source/public domain content, so any contribution by AI tools must be considered released under public domain. It does not get simpler than that, and current US copyright law has already indicated that any AI created works are not eligible for copyright. So disclosing it as partially or completely AI generated = Public Domain. If at some point someone can prove that code created by a tool was pilfered from a OSS project, then you can re-disclose it

Re:

[Ostracus]

Might want to run that by a lawyer because the US isn't the world even if it acts like it.

Re:

[Valgrus Thunderaxe]

Those AI tools were trained on open source/public domain content,

So was I, but I get to chose whatever license I want for the software I write.

Re:

[phantomfive]

That is how it's been, Those AI tools were trained on open source/public domain content, so any contribution by AI tools must be considered released under public domain. It does not get simpler than that, and current US copyright law has already indicated that any AI created works are not eligible for copyright

That's not the question.

The question is whether the AI-produced code is a derivative of existing code, and the answer is still not resolved.

In some cases, the answer is a clear YES, because the code is a direct copy of something written by someone else. If something like that ends up in the kernel, it will have to be removed when someone notices.

Re:

[Meneth]

Linux has a number of rules for licences [github.com] in the kernel source tree. "Public domain" code is not mentioned. It seems implied that every contribution has some well-defined license.

Re:

[HiThere]

Yes, but if you take several pieces of public domain code and create a new composition with them, that new composition is copyright. (I don't think you can even avoid having a copyright, though you can have a license that is essentially the same as public domain.)

Well...now I suppose you can avoid a copyright by feeding it into an AI and then having the AI regurgitate it. Or, if I've understood the news stories correctly, by claiming that an AI wrote it rather than "wrote the basic parts which you later a

AI code

[gary s]

Who is requesting the AI to generate code? that should be the submitter, not an AI bot. Who will be responsible for long term management of the code when it has issues.

Re:

[Teun]

Indeed, as long as a human bears the end responsibility there is no problem.

yes*

[ClueHammer]

(*) If it has been checked, and adds something positive to the project. Sure add it. If its unchecked AI slop the poster needs to kneed in the happy sacks till they learn to do better.

Re:

[HiThere]

That's clearly the ideal approach. The problem is that you're likely to end up drowning in submissions.

I wonder..

[Junta]

include performance improvements that indirectly support AI applications,

I wonder how the developers feel about that characterization. Now all performance improvements are subject to being attributed to motivated by AI for marketing points...

AI vs GPL?

[Predius]

So, in the US machine generated content isn't eligible for copyright. What does that do to a project under GPL or other licenses? Does the license become unenforceable once a certain threshold of non-copyrightable content is in the project?

Re:

[HiThere]

IIRC, in music copyright cases one bar of sufficiently similar notes is enough to justify suing for copyright infringement. And a composition of names is enough to justify the grant of copyright.

Law being what it is, that doesn't really prove anything, but it strongly suggests that a "novel combination" of code should be copyrightable even if all the pieces are public domain.

Re: AI vs GPL?

[OrangeTide]

You would have to replace all the human created code before you could even begin to challenge the copyright. Even then, it may not be possible to hijack a very long running project's copyright status (never has been done before, as far as I know)

Theoretically you could extract the bits that cannot be held under copyright and do whatever you want. But only in theory, because you still might have a court case on your hands and victory is not guaranteed. Civil suits can get really messy, really fast.

Re:

[allo]

Two factors come into play. First, the copyright problem you think of only applies to completely generated code without human changes. That is unlikely to fit into the kernel. And second you can add as much public domain code to a GPL project as you want. The only issue is, that the public domain code can be reused without any license, attribution, whatever. But your GPL code stays GPL. Code licenses are no majority voting, but need to be compatible and public domain -> GPL is compatible (the other way r

AI should fork it

[snowshovelboy]

AI should fork it and apply its own changes. And then a different AI should code review it. Another AI can test it, etc. How should the AI handle human generated contributions to its fork?

Re:

[Quantum gravity]

Your logic is inconclusive. You require a reboot.

use ai

[groobly]

Simple. Use AI to evaluate AI contributions. Preferably not the same AI, but in my meager experience the same AI (a) is good at finding its own mistakes, and (b) randomly appears to choose different solutions to the same problem.

Two words

[Bu11etmagnet]

/dev/null

ZDNET warns

[allo]

I doubt Linus Torvalds cares about ZDNET warnings.

Send to /dev/null

[devslash0]

That's about it.

Simply no.

[hebertrich]

There's too many issues with copyright and AI companies to take a risk with something as fundamental and important as the linux kernel .
Simply put ? No. AI contribs should simply be dismissed.

Denile

[BrendaEM]

The crap must end. We need reins on AI proliferation.