Linus Torvalds Rejects an AI Ban as Linux Sets Rules for Assisted Code
Linus Torvalds’ position on AI has changed in one important respect. In an interview published on October 17, 2024, the Linux creator treated practical value as something that still had to emerge from the hype; a July 2026 account of the kernel mailing-list exchange documents his refusal to let opposition to AI determine who may contribute to Linux.
The shift does not amount to unconditional trust in generated code. Current Linux kernel guidance for AI coding assistants keeps licensing checks, review, Developer Certificate of Origin certification and final responsibility with a human submitter.
What the 2024 criticism was aimed at
The 2024 position combined interest in the technology with deep skepticism about the industry surrounding it. Torvalds expected AI to matter, but considered demonstrations and promotion insufficient evidence of durable usefulness in ordinary workloads.
That distinction is central to understanding the later change. The criticism was directed at the gap between expansive claims and demonstrated utility, rather than at every application of machine learning or every developer who experimented with an AI tool.
The widely repeated description of the industry as “90 percent marketing and ten percent reality” was an opinion about that imbalance. It was not a measurement of model accuracy, a market forecast or a benchmark comparing AI systems. Treating the phrase as technical data gives it a precision it never had.
The practical test embedded in the criticism was more demanding than whether a model could produce an impressive answer. Useful engineering software must operate inside established processes, produce work that can be checked and avoid leaving maintainers with more verification labor than the tool saves.
Linux accommodates AI assistance without transferring responsibility
The kernel rules apply the existing contribution process to machine-assisted work instead of creating an easier route into the project. Contributions must remain compatible with the kernel’s GPL-2.0-only licensing requirements and use appropriate SPDX identifiers.
An AI agent must not add a Signed-off-by tag. That certification belongs to a person who has reviewed the generated code, checked the licensing position and accepted responsibility for the contribution under the Developer Certificate of Origin.
The documentation also calls for an Assisted-by tag identifying the agent and model version, with specialized analysis tools included when relevant. This attribution records the tool’s involvement, but it does not make the tool an author capable of certifying provenance or accepting responsibility.
The distinction is more substantial than a labeling convention. A human contributor cannot treat a plausible patch as finished merely because an assistant generated it: the change still has to build, survive testing, fit the surrounding code and withstand review by the maintainers responsible for that part of the kernel.
The guidance is therefore neither an endorsement of arbitrary generated output nor a prohibition on AI-supported development. It recognizes assistance as part of a possible workflow while preserving the same accountability chain used for other contributions.
What changed by July 2026
By July 2026, the unresolved question was no longer whether useful AI applications might eventually appear. The immediate dispute concerned whether hostility toward the technology should prevent other kernel developers from using it, including in systems intended to assist patch review.
Torvalds drew a firm boundary around that question: Linux would not become an anti-AI project, and contributors who rejected that direction retained the open-source options of maintaining a fork or leaving. He also framed AI as a tool whose value should be judged by whether it improves technical work.
This does not mean that every AI-assisted patch deserves acceptance or that maintainers must use such systems themselves. Tool choice and patch quality remain separate questions. A contribution can disclose AI assistance correctly and still be rejected because it is wrong, untested, legally unclear, unnecessary or too costly to maintain.
The newer position is compatible with much of the older skepticism. Marketing claims can remain exaggerated even after particular tools become useful. Recognizing a productive use does not require accepting an industry’s forecasts, economics or claims about automation.
The enduring issue is verification cost
Kernel development exposes a basic asymmetry in generative software tools. Producing a patch, review comment or bug report may be cheap, while checking it can require reproducing behavior, studying surrounding code, testing hardware-specific paths and assessing possible regressions.
When the person submitting automated output does not perform that work, the labor has not disappeared. It has been transferred to maintainers, who must separate genuine findings from confident but invalid suggestions. Enough low-quality submissions can make a productive tool harmful at the project level even when individual outputs occasionally look convincing.
Human certification addresses that problem by attaching generated work to someone who can answer questions, revise the patch and defend its provenance. The Assisted-by tag provides transparency; the human Signed-off-by tag establishes accountability. Neither substitutes for technical review.
For software creators, that is the most consequential update to the 2024 controversy. Torvalds has moved beyond simply tuning out AI promotion, but Linux has not replaced engineering judgment with model output. AI-assisted work can enter the process; trust still has to be earned through evidence, review and an identifiable person willing to stand behind the result.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.