Linux Maintainers Can Now Submit AI-Generated Code Without Declaring It (xda-developers.com) 86
Linux maintainers "have ruled that people no longer have to declare their LLM usage," reports XDA Developers, "as Torvalds himself calls a ban on AI 'utterly stupid.'"
The first part of this story comes via a message sent to the Linux mailing list. It announces one of the major changes decided during the Maintainers Summit on October 8, 2026, in Prague, Czechia, where submissions to the kernel no longer need to use the "Assisted-by" tag... meaning that people can submit AI-generated code without needing to declare it as such.
In some correspondence regarding concern about this new disclosure removal, the creator of Linux, Linus Torvalds, made his thoughts known... [Torvalds] compares an AI ban to getting rid of text editors, as doing so would prove that someone could write a good commit message without making any mistakes. He then rounds off his argument by asking, "Do you not see how UTTERLY STUPID those kinds of rules would be?" and stating that he simply expects the people working on the Linux kernel to "understand what they are doing."
Here's the last line of Torvalds' post. "Quite honestly, in my experience, the people I respect and call intelligent are not the ones that shy away from using new and interesting tools."
In a related story, Torvalds says he's used AI to help him understand patches that were hard to follow. And this week Torvalds even posted an LLM-assisted patch, reports Tom's Hardware: The development branch purports to fix an old issue with massive cache contention when many CPUs are reading the same file — a scenario potentially encountered when using shared libraries, databases, and web servers. Several developers had already made attempts over the years at fixing the issue, all with varied results, none quite ideal as an all-encompassing solution... He's now posted his solution in a development branch, and claims that he's seen a 23x performance improvement in a dedicated synthetic benchmark, though he notes that actual real-world impact will be far less significant, as "no real load" will witness those figures.
Torvalds used Anthropic's Claude Code, reports The New Stack, and "said the assistant wrote the implementation, tests, and initial documentation, and reviewed the code under his direction. He ran the benchmarks himself, despite having previously criticized developers for claiming AI writes 99% of their code." Tuesday Torvalds submitted a 10-patch series.
"Reviewers have since identified a race condition and raised concerns about memory reclamation," adds The New Stack. "Torvalds has handed further development to Kiryl Shutsemau, a Linux kernel developer who previously worked on the same bottleneck."
In some correspondence regarding concern about this new disclosure removal, the creator of Linux, Linus Torvalds, made his thoughts known... [Torvalds] compares an AI ban to getting rid of text editors, as doing so would prove that someone could write a good commit message without making any mistakes. He then rounds off his argument by asking, "Do you not see how UTTERLY STUPID those kinds of rules would be?" and stating that he simply expects the people working on the Linux kernel to "understand what they are doing."
Here's the last line of Torvalds' post. "Quite honestly, in my experience, the people I respect and call intelligent are not the ones that shy away from using new and interesting tools."
In a related story, Torvalds says he's used AI to help him understand patches that were hard to follow. And this week Torvalds even posted an LLM-assisted patch, reports Tom's Hardware: The development branch purports to fix an old issue with massive cache contention when many CPUs are reading the same file — a scenario potentially encountered when using shared libraries, databases, and web servers. Several developers had already made attempts over the years at fixing the issue, all with varied results, none quite ideal as an all-encompassing solution... He's now posted his solution in a development branch, and claims that he's seen a 23x performance improvement in a dedicated synthetic benchmark, though he notes that actual real-world impact will be far less significant, as "no real load" will witness those figures.
Torvalds used Anthropic's Claude Code, reports The New Stack, and "said the assistant wrote the implementation, tests, and initial documentation, and reviewed the code under his direction. He ran the benchmarks himself, despite having previously criticized developers for claiming AI writes 99% of their code." Tuesday Torvalds submitted a 10-patch series.
"Reviewers have since identified a race condition and raised concerns about memory reclamation," adds The New Stack. "Torvalds has handed further development to Kiryl Shutsemau, a Linux kernel developer who previously worked on the same bottleneck."
Re:Damn. (Score:5, Interesting)
Do you think for some reason that the changes won't be tested?
By the looks of things Linus did do enough testing to satisfy himself there was progress made.
Linux itself gets huge numbers of CPU cycles run by it each day.
I think his approach is pragmatic.
I personally don't use LLMs for anything because if I have to keep checking the results for factual fabrications aka hallucinations then I'm wasting my time, but I don't think they're going to go away either so we'll see how it works out.
Yeah we kind of think that (Score:4, Insightful)
Re: (Score:2)
Re: (Score:2)
Do you think Torvalds reads studies, or do you think he have hands on experience with development and code review?
Re: (Score:1)
Every tool has the potential to give you productivity gains, and every tool, if used incorrectly, will more than likely yield the opposite. That isn't anything new.
Re: Yeah we kind of think that (Score:2)
Re: (Score:1)
I'm not in the habit of interpreting fiction as a sort of prophecy.
Re: (Score:2)
And every tool has a limited range of things where it can actually give you those productivity gains (if used correctly). It looks more and more like production code generation is NOT an area where you can get those improvements from an LLM.
Re: Yeah we kind of think that (Score:2)
The speed with which AI is improving, I am not sure that a study of even six month in age is that valuable.
Of course code review has become a bottleneck. But that also use AI. Indeed, I think, anyone who is writing code without an AI review is a problem.
We are rapidly moving to a situation where once the need is clearly identified, the code will just appear. Time to move up an abstraction level.
Re: (Score:2)
Yes. There have been several attempts to automate software generation. When I studied CS, the 5GL project had just failed resoundingly. The idea was to write specs in constraint logic and then have the machine write code. What made it fail? Turns out writing a full spec is a lot more effort and requires significantly better qualifications and experience than writing the code directly. The idea of getting code automatically from a spec has since failed again with other approaches, essentially due to the same
Re: (Score:2)
The problem is not on the tech side. The problem is on the legal side. The problem is that LLM code does not get new copyright. If nobody was copied close enough, LLM code does not qualify for copyright and you cannot slap a license on it. But if it is close enough to something in its training data, it will very likely be found to be still unter the original copyright, license and ownership. And that is a legal nightmare.
Yes, this nuke may need a few years to detonate. But anybody that followed the SCO laws
Re: (Score:2)
BSD is always here as an alternative, but don't be too hasty to jump boats.
New and interesting tools? (Score:2)
I'm sure a chainsaw is interesting but you wouldnt use it to slice your bread. It has to be the right tool and while current AI can write code it often has subtle bugs which the AI user wont have a clue about but might have avoided if writing the code himself.
Re: New and interesting tools? (Score:2)
Wizards have been using generative AI assistance since the mid 2000s. Finance professionals have been automating work since the 1960s. Folks have been using compilers instead of writing machine code for decades.
Outsourcing the non-value adding parts of the work to automation wonâ(TM)t hurt the folks who put time into learning and knowing what theyâ(TM)re doing. Where this could be impactful is for folks who donâ(TM)t know what theyâ(TM)re doing and are unwilling to put in the effort to g
Re: (Score:2)
"Outsourcing the non-value adding parts of the work to automation wonâ(TM)t hurt the folks who put time into learning and knowing what theyâ(TM)re doing."
What parts of coding are "non-balue adding parts"?
"Where this could be impactful is for folks who donâ(TM)t know what theyâ(TM)re doing and are unwilling to put in the effort to get good."
Like you as reflected in this post?
Re: New and interesting tools? (Score:1)
In my experience, most of the value in good software development is creates before a line or code is written. Itâ(TM)s in the specifications, getting the business logic right, asking the right questions, designing a good architecture, etc. which AI is not even close to becoming good at.
Even modestly competent developers should not be threatened by AI. But even the best coders are replaceable as has been seen in the past few years.
Re: New and interesting tools? (Score:2)
Generative ai didnt exist outside of theory books in the mid 2000s. HTH
Re: (Score:2)
Re: New and interesting tools? (Score:2)
I am not happy with the death of coding as a occupation and hobby. But using a decent model it can write some pretty hardy tests into a reusable harness and often delivers better than average code when it comes to maintainability. Having a comprehensive test suite is like a few lines of prompting. Do that over the next 100 commits and there are a lot few bugs and the AI itself flails around less making changes on your sturdy foundation.
You can't watermark text (Score:2)
Banning AI, especially in source code, would be a nightmare to track.
There's no way to see if a file in ascii was written by AI, or written by a human.
There's going to be a lot of people condemning him for this, but there's a world of difference between using AI to help your work, research and do the tedious stuff, and using AI to vibe code without knowing what you're doing.
The best use for AI is doing drudge work and finding needles in haystacks.
The genie isn't going back in the bottle no matter how much a
Re: (Score:3)
"There's no way to see if a file in ascii was written by AI, or written by a human."
Yeah that's the reason. Who cares about merit when we can be concerned about the tools used. No doubt this is the key conversation among kernel developers!
Re: You can't watermark text (Score:2)
Finger printing text already exists and is currently deployed. But it only helps you with the models that do finger printing and you can intentionally reverse it with a pretty trivial local model.
Re: (Score:2)
So do you think
- counter = size;
+ counter = size +1;
is LLM generated or human generated? Do you think there are many different ways to write this fix?
Re: (Score:2)
Actually you can watermark text: https://www.anthropic.com/news... [anthropic.com]
Re: You can't watermark text (Score:2)
Actually you can watermark sufficiently long samples of text, but the watermarking won't survive any significant and worthy editing process.
Well of course... (Score:1)
Its not the same torvalds using an LLM than silly me with my 99 IQ calling myself kernel vibecoder, gatekeeper slayer....
Re: (Score:2)
His LLM code ended up sucking even though he tried to test it, so it might actually level the playing field.
Not in the way most people would hope for, but still.
What a monumental historic relevant mistake (Score:3)
Decades after inventing one of the most influental and important operating systems he destroys it within a couple oft days. Like burning rome to the ground.
AI is about to destroy society as we know it. Not in a good way, like disrupting things for the better, which is sometimes necessary.
No, people already start giving away their independence and their ability to learn new things and 'lets simply ask ChatGPT'. The very thing that made us go forward, the curiosity to learn new things, the will to endure the frustration to fail until you finally get it (whatever 'it' is) is not slowly but within months erased. 'Just ASK ChatGPT' is the new mantra. Don't know how to code? Use Claude. Give decision making to an AI and everything will be fine. Great.
My first distro was SuSE Linux 4.4.1, the last one I used despite my Raspis was Devuan (since I am pretty skeptical ofe systemD, yes I know the Raspis use it). So today marks a sad day.
Re: (Score:1)
Torvolds did not invent linux, he copied it from existing work. And linux would be nothing without AT&T and their lawsuits, linux succeeded by accident. Kids these days.
"No, people already start giving away their independence and their ability to learn new things and 'lets simply ask ChatGPT'. "
Never ONCE heard someone say that, no doubt someone has but I deal with people post-puberty.
Re: What a monumental historic relevant mistake (Score:2)
And here I thought Linux isn't Unix. But now the grognards tell me it is just a copy that didn't contribute anything new on its own.
I would like to go on record that operating systems existed before Unix. Many of them had file handles and process isolation too. Most of them were much more complicated than the first three versions of Unix. Designed for nicer and more profitable hardware.
Re: (Score:2)
And here I thought Linux isn't Unix. But now the grognards tell me it is just a copy that didn't contribute anything new on its own.
Well, it wasn't Unix. But it was inspired by Minix, a widely-taught teaching operating system that Torvalds decided to make into a thing you could use every day.
Re: (Score:2)
inspired != copied
AI models likely have back doors (Score:2)
In seeing posts from other people saying they the code they have written is above average in quality. With the amount of code generated most people are not reading it, and to truly understand it you often have to piece it together yourself.
Most projects nowadays people are not really reading the code, and I have a hard time not thinking that it won't happen to Linux development.
Models like deep seek most likely have Chinese backdoors. The kind where specific phrases trigger the black logic. The US models
Pragmatic Approach, Unmentioned Caveat (Score:5, Interesting)
As usual, Torvalds' philosophy is pragmatic. I more or less agree with the gist: Look at the code, not where it came from.
Also as usual, certain details have been swept under the rug—chiefly that Torvalds' only significant vibe-coded contribution to the kernel thusfar was a ten-patch modification to page-caching that was promptly rescinded when other maintainers pointed out a race condition hidden in Claude Code's verbose output. [thenewstack.io] Torvalds ultimately dumped the mess onto human coder Kiryl Shutsemau's lap.
This raises an important, obvious question.
The part I don't get (Score:3)
Re: (Score:2)
But if you ignore the results he seems to think it's the bee's knees.
As they say, what's remarkable isn't that the dog plays a good game of poker; it's that the dog is able to play poker at all.
Re: The part I don't get (Score:2)
Indeed, it's remarkable, but you wouldn't hand the dog your money and send it to the WSOP.
Re: (Score:2)
The AI code I've seen and generated isn't nearly as good as human code, but that doesn't mean it is worthless. If I'm creating a script or little app for hobby stuff, to do one very specific task, I can either spend my own time on it and probably over-engineer the thing, or I can get AI to crap out something that basically works. So I decide based on if I think I will enjoy writing the code, or if I just want to get on with the other parts of the hobby and remove one manual process that is a slog.
Companies
Re: (Score:2)
Linus Torvalds [thenewstack.io]: “AI is a great new tool, but it’s a tool, and when I see people saying, ‘Hey, 99% of our code is written by AI,’ I literally get angry, because those same people — I can pretty much guarantee — that 100% of their code is written by compilers. But they never say that.”
Re: The part I don't get (Score:2)
Re: (Score:2)
Re: (Score:2)
How dare you identify the failure being claimed as success!
Re: (Score:2)
Perish the thought that the lead kernel maintainer would collaborate with other developers
--
“
Grabs the popcorn (Score:2)
I can't wait to see the comments on this. The few are already posted are providing good entertainment.
Toothless rules or witch hunt - none is good (Score:2)
As much as I dislike Megacorp-powered AI for coding... these declarations of usage are either unenforceable, or they will create too much friction and witch hunts (in already a ... temperamental community). So, it's a pragmatic decision, the way I see it. But the contribution volume needs to be seriously controlled, otherwise they'll validate Antropic's Scion approach.
Re: (Score:1)
I agree with your conclusion, but the problem isn't that the community is "temperamental." It's that programmers are generally emotionally immature and have low general intelligence. Including Linus. Absolutely a software genius, but every time he tries to argue a controversial point he makes an apples to oranges comparison and then uses it as a straw man, while thumping the table, calling people stupid, and declaring absolute unquestionable victory.
There's just no way for any of them to be good at enforcin
Copyright (Score:5, Interesting)
Re: (Score:2)
No, that's not the biggest problem if it is a problem at all.
"... but I doubt that describes "Claude"."
Or even a viable LLM.
Re:Copyright (Score:4, Interesting)
Indeed. LLMs are like a cryptocurrency tumbler for intellectual property. It is a metaphysical certitude that, sooner or later, a major legal challenge is going to happen with Linux and/or other open source software collections from some rent seeking copyright holder(s) due to AI contributions. Linus is taking a real risk here.
May you live in interesting times. Whatever outcome emerges from the forthcoming years long legal processes will set profound precedents.
I predict this won't actually start with Linux, however. It will be Adobe. There is a cottage industry of Photoshop clones being created with LLMs. Adobe is going to take someone to court over this.
Re: (Score:2)
Re: (Score:2)
The "problem" is that the output may be public domain. But if you add public domain code to "viral" GPL code you end up with a GPL product. So who cares?
Re: (Score:2)
Re: (Score:2)
I see your argument for someone "one-shotting" a tetris game as benchmark (even though this shows much diversity for most models). But if you load a Linux file and then ask the LLM to check all memory allocation and let it implement the suggested safeguards you can be pretty sure that this code never existed before. Its the different between asking for "Write me a story", asking for a story by explaining in two strong paragraphs what it should be about, and asking an LLM to help you to refine your story so
Re:Copyright & Declaration (Score:2)
Yes. The question whether people are allowed to use AI for every aspect of kernel development is different to the question of transparently declaring use of outside material (the model). I understand the reasoning for deciding on the former, but not for the latter. Now we will no way to tell.
The comparison to compilers, checkers and editors is not a good one for copyright and intellectual attribution.
Re: (Score:2)
Exactly. Linus is creating a potentially catastrophic legal problem here. And he does so after the GCC folks have clearly explained the problem. Stupid. Or maybe he is getting old and does not care that much anymore.
Just Wait⦠(Score:1)
till he gets flooded with slop pull requests..
calls a ban on AI 'utterly stupid (Score:2)
I agree, nonetheless I value my YouTube AI blocker extension very dearly.
Counter point (Score:2)
(a) What's the harm in labeling AI generated code as such, and
(b) How difficult is it to do so, especially as the code is generated?
Knowing who and/or what created/edited the code can probably be helpful later.
Re:Counter point (Score:4, Insightful)
Re: (Score:2)
Think about it down the road after ten more patches. Can you really say the line is AI code, when much code around it was human changed? If the line would not have been there the human would have added it as human line. So why label it AI, when there is no "fair" label whose line it is? Git blame is a tool, but no good source of truth for copyright or attribution.
Re: (Score:1)
it has been decided I think that tagging language that is not unique to the instance is counterproductive. the argument that AI agents copy code will be killed forever, and by the way if they do or not is irrelevant since most people do in my opinion, once the argument is killed the only people left to oppose it are the people using it and telling others not to.
He's right (Score:2)
Linus has a radical stance on these things. Just like he says "No special treatment for security bugs, because we fix ALL bugs", he now says "No special treatment for AI commits, we scrutinize ALL commits". Who cares if your commit is slop because the AI is stupid or because you are stupid? It is merged, when it is good code, and Torvalds shouts at you if it is bad code.
Re: (Score:2)
The problem is not with code quality. The problem is with copyright and license. They are currently falling into a potentially deadly trap, same as anybody that accepts LLM generated code.
Re: (Score:2)
I bet he thought about that more than the Slashdot commenters and came to some conclusion ...
Re: (Score:2)
If you depend on everybody here being incompetent, you are dead wrong.
Re: (Score:2)
Everybody but you, of course!
Re: (Score:2)
An AdHominem like this says things about you, but not about me.
Re: (Score:2)
I said everyone *but* you is wrong. So it doesn't say anything about you.
A Linux Self-Inflicted Wound (Score:2)
Beware of knee-jerk reactions (Score:3)
Caution is required, as it is with any submission to the Linux kernel, but I think it's sensible for Linus to be open to the idea (albeit with his guns-ablazing turn of phrase).
At the other extreme, we need to be careful of the knee-jerk reaction of writing off all AI generated code as "slop". Sometimes this can be no more than a sort of "signalling". If I label code as "slop" I am really just letting you know that I am a programmer at a level that can discern such things... even if I am not. (Since feeling I need to let you know betrays an insecurity).
In my limited experience, AI can generate good code, particularly for small, well defined cases. And even when the first draft is not good enough, you can refine the requirements and get it to improve the code. Just as you might with a human.
Re: (Score:2)
At very best the LLM output is not actually copyrightable but it is being distributed as if it were GPLv2 and it's fraudulent to distribute on that basis. At worst it actually carries copyright from contributors and it will have to be removed and replaced later. Any project allowing AI code contributions is creating a minefield. It is just an absolute dogshit plan to accept them for this reason alone and I absolutely hate that this is happening to Linux.
Re: (Score:2)
Exactly. If you allow LLM generated code in your project, you may lose it completely. Suddenly it may be under a different license, or, worse, several different ones. And your ownership and original license may be gone. The GCC folks understand that catastrophic risk. It looks like Linus does not or does not want to deal with it. That does not work and is an exceptionally bad idea.
The absolutely very least and lowest level of caution would be mandatory tagging of all LLM written code. Because with that you
Re: (Score:2)
At very best the LLM output is not actually copyrightable but it is being distributed as if it were GPLv2 and it's fraudulent to distribute on that basis. At worst it actually carries copyright from contributors and it will have to be removed and replaced later. Any project allowing AI code contributions is creating a minefield. It is just an absolute dogshit plan to accept them for this reason alone and I absolutely hate that this is happening to Linux.
That's an interesting point. I can see how companies might register this as a risk, and avoid it on that basis, even if the legal conclusions have not yet arrived.
To clarify, when you describe LLM output as copyrightable, are you implying the copyright holder is the LLM provider, or the owners of the original training data?
In discussions of copyright on this topic, the theme is usually a complaint about LLMs being trained on data created by humans, without the permission of those humans. They might have a c
Re: Beware of knee-jerk reactions (Score:2)
"To clarify, when you describe LLM output as copyrightable, are you implying the copyright holder is the LLM provider, or the owners of the original training data?"
To clarify, nobody knows yet! But I do mean the owners of the original data. It's not actually been decided whether their copyright can apply to code produced by LLM which was trained on it, and until it has been, at the VERY MINIMUM the AI code should be required to be labeled. Not doing so is willfully pretending this is not an unsettled legal
The problem with automation is (Score:2)
AI has patterns of growing pull requests. While it would be nice if Linux can buck the trend, I do not believe you can force human beings to verify code they have not written. Part of the iterative process of coding is verifying things work the way you think they do as you add it.
There is also the niceness of knowing what you are looking at before getting into a a PR. When you know you are possibly looking at slop before looking at it, the pain if being stabbed in the back is much better. Though I guess
What does that have to do with it? (Score:2)
This story makes no sense. They decided people no longer have to declare they're using AI, and then justify it by saying banning AI would be stupid? What do those have to do with each other? Banning AI is totally different from requiring people to say they've used it. Objecting to the first has nothing to do with objecting to the second.
Yes, I read the original story. It's just as confused as the summary. Does someone know what actually happened?
So, proper attribution doesn't matter? (Score:4, Interesting)
This makes little sense to me. It's not a huge burden to declare you used LLM tools. It is a acknowledgement that the code was generated from tools that have used sources that are not attributed nor provide primary sources for the derivative works they create.
My bias is that all the major LLMs have abused the notion of fair use to an completely unreasonable degree. It's a stretch from allowing academic teaching and research to exist to making all the monies for some tech startups.
It would be different if these tools were indeed freely available for non-commercial and research purposes, but that's literally too expensive to do.
Linux, a project whose success is directly tied to the GPL-2 license and the protections it afforded is now just shrugging it's shoulders and saying why even bother just noting that code came from a tool that inhaled everything out there, because it's useful?
That's just lazy.
*BSD, here I come (Score:2)
It was good for as long as it lasted.
No good can come from using AI on such a big code base as the Linux kernel.
LLM ~ Compiler? (Score:1)
I'll steer clear of the code attribution/copyright questions. That's too complex for my brain. I'll also steer clear of the fact that one should understand the code they submit. That's obvious, surely.
What I'm primarily confounded by is Linus's "AI is a programming tool like a compiler is" analogy. I first heard him make it here [youtu.be] (the video starts where his remark does). His recent mailing list email seems to be an extension of that line of thought of his, together with the "editors are programming tools too
Re: (Score:2)
An LLM is not a compiler or a valid step between specification and code, unless you put a lot of safeguards and extra effort in place. First, an LLM comes without specification of what it will do. The Language specification a compiler is based on is critical for it to be useful at all. Otherwise you have no reliable idea what the compiler will do with some input and things may change without warning. For an LLM, this is even worse: Things may change between one run and the next and things may change an enti
That is a really bad idea (Score:2)
Not because of patch quality, the quality gates on the kernel are fine. It is a really bad idea because of copyright. Depending on how the copyright battle in the courts goes, this may seriously taint the kernel and removal of the tag may make cleanup impossible.
Shows that Linux, unfortunately, does not understand the legal angle.