The Linux kernel spent most of a year asking developers to confess when a machine helped write their patch, and it has now stopped asking. The rule, introduced in January, is being removed, partly because not everyone followed it and partly because Linus Torvalds thinks it was aimed at the wrong target.
What the rule said, and why it is going
Under the policy, contributors were supposed to add an Assisted-by tag to a commit whenever they used any advanced coding tool. It was a disclosure rule, simple on paper.
On 9 October, Sasha Levin, a developer and Nvidia engineer, proposed deleting it. He pointed to a consensus reached at the Maintainers Summit, so this was less a rebellion than the paperwork catching up with a decision already made in the room.
Jeff Layton of Red Hat then added two practical objections:
- Patchy compliance. Not everybody follows the policy.
- Free advertising. In his words, the tag works as unpaid promotion for proprietary LLM companies.
The first point is the awkward one for any honor-system rule. If only the conscientious comply, the tag measures conscientiousness, not AI use. The second is a nice irony: a transparency measure that quietly turned the kernel's commit history into a billboard.
The objection, and Torvalds' reply
Not everyone agreed. Laurent Pinchart, founder of Ideas on Board, argued against removal. His concern was that without some other way to disclose AI involvement, it becomes harder to screen out low-effort or dishonest submissions.
Torvalds answered bluntly. As reported, he asked whether Pinchart could not see how silly such rules are, then suggested that Pinchart stop using compilers to submit code if he wanted to prove he understood it. The jab makes a point. A compiler is also a tool that turns human intent into something the machine runs, and nobody tags commits with the compiler's name.
His underlying position, as the report relays it, is that expectations should rest on what the developer understands about the change, not on which tools were involved. A patch is judged on whether its author can explain and stand behind it. How it was typed is beside the point.
The quotes here come from a Russian-language report and have been retranslated, so the wording may differ from the original mailing-list exchange. The substance is consistent: Torvalds wanted accountability attached to people, not to toolchains.
Torvalds is not an AI skeptic
This was not a surprise given what he said earlier at Open Source Summit Europe. There, Torvalds admitted he now enjoys using AI for tasks he is not good at. His example was writing Java for a guitar pedal project of his own.
He called vibe coding a wonderful way to find joy in programming. He also added a caveat: when you build something real and important, you need to be careful. That is a reasonably consistent philosophy. Use the tool freely for a hobby, and hold the work to a higher standard when other people depend on it.
The kernel's mixed experience with AI
The upside is already visible. AI tools are finding real vulnerabilities in the kernel, including in drivers that nobody had touched for 20 years. Two decades of dust, it turns out, is a fine place for bugs to sit comfortably.
The downside lands on maintainers. According to the report, three quarters of the talks at the summit were about reducing the stress AI creates for them. Developers joke that the bots are now talking to the bots, which is funny until you remember that a human maintainer is still the one reading the thread.
So the kernel's situation is both better and more tiring. Tools find old flaws, and the same tools generate more material for people to review.
Other projects are heading the other way
The kernel's relaxed stance is not universal. Two other open source projects have gone in the opposite direction:
| Project | Position |
|---|---|
| System76 | Banned AI content in its COSMIC environment |
| NetBSD | Has banned AI-written code since 2024 |
NetBSD's ban seems to have done it no harm with its community. The project collected $120,000 in donations against a goal of $50,000. A firm no is, at least, a clear message to rally around.
What this actually changes
In practice, the kernel is trading a disclosure ritual for a responsibility principle. The tag is gone, but the expectation that a submitter understands their patch remains, and that expectation is the one Torvalds cares about.
Pinchart's worry has not been answered, only overruled. Without a label, reviewers will need to spot weak submissions the old-fashioned way, by reading them and asking questions. That is more work for the people already reported to be stretched thin by AI-generated volume.
There is also a fair question about whether a rule that only the honest obey was ever doing its job. Layton's observation suggests it was not. Removing it at least ends the pretense.
The report does not say whether any replacement mechanism is planned, so that remains open. For now, the kernel's answer to the age of AI-assisted code is the same as its answer to every earlier tool: show that you know what you sent.