On October 13, 2026, Linus Torvalds, in an email to LKML, responding to Laurent Pinchart in a thread about the patch 'docs: remove Assisted-by references,' stated that he does not understand the concept of 'don't use AI for X' and refused to introduce rules banning AI tools when working on the kernel. With a sarcastic 'new rule,' he suggested banning compilers, since if a developer cannot write code in assembly, they do not understand what the machine is doing. The maintainer's key thesis: contributors are expected to understand their own actions, and this understanding 'has very little to do with what tools they use.' The statement came against the backdrop of Torvalds himself preparing a perf page cache fix through Claude Code, although his branch did not survive review — a race condition was found there.

image

What happened

The email appeared on October 13, 2026, in the LKML mailing list: Linus Torvalds replied to Laurent Pinchart in a thread about the patch 'docs: remove Assisted-by references.' Torvalds wrote that the formula 'don't use AI for X' makes no sense to him, and he will not introduce bans on using AI when preparing contributions to the kernel. In response to critics, he sarcastically 'established a new rule': compilers are banned in submissions because a developer who cannot write code in assembly does not understand what the machine is doing, and further by this logic — 'you had better not use an editor to write that commit message either.' He summed up his position as expectations, not tools: 'The kind of expectations I have is that people do their best and I expect people to understand what they are doing. But that understanding has very little to do with what tools they use,' noting that people he respects do not hide from new technologies. This email in the thread and the maintainer's personal position, not a change in official patch acceptance rules.

Context

The background of the discussion is a dispute over marking AI contributions: on October 9, 2026, Sasha Levin proposed a patch removing the Assisted-by trailer from kernel documentation, and the thread around it raised the question of whether AI tools should be restricted at all. In parallel, this topic was already being tested in practice: Torvalds himself was preparing a perf page cache fix in the noref-filemap-read-with-llm branch using Claude Code, meaning assisted code went through the standard Linux review process as a regular patch. The branch did not reach a merge — it was the reviewers, not the author or the agent, who found the race condition, and further work on the fix was handed over to Kiryl Shutsemau. The claimed throughput increase of up to 23x for this fix exists only in a synthetic benchmark without a described methodology, so for now it is a hypothesis, not a measured result. The mechanics themselves are telling: the quality of the diff was caught by regular kernel review, not some separate AI-specific control.

Why this matters for the industry

The main gatekeeper of the largest open-source project publicly established a norm: the policy around AI tools relies on developer competence and mandatory review, not on tool bans. For teams building product pipelines, this is a ready-made workflow template 'agent prepares a draft of changes — strict review and tests,' and it has become noticeably easier to reference the Linux maintainer's public position when legitimizing AI agents in an internal codebase. At the same time, the fresh kernel case outlines the boundaries of agentic coding: concurrent code remains a weak point of generation, an error in this class was caught by reviewers, not a tool, so shifting investments from pure generation to verification — static analysis, auto-generation of tests, race detectors in CI — is now a more justified move. If the Assisted-by trailer really disappears from the documentation, the ecosystem will lose marked data on AI code provenance, and it will become impossible to retrospectively assess the frequency of defects in assisted diffs. For AI coding vendors, the kernel's position is a strong argument in sales to conservative segments, but only with an honest presentation of the case without unvalidated benchmarks.

Why this matters for users

The material is easy to verify from primary sources: Torvalds's email is fully open on lore.kernel.org, and the discussion on Hacker News at the time of recording had 7 points and 4 comments. Practical takeaway for developers: using an AI assistant for both code and commit text is not a violation — the error begins where the author does not understand their own diff, so a working guideline sounds like 'generate boldly, but only what you can explain and defend in review.' For those trying agentic tools in their pipelines, the page cache case gives a clear lesson on the test plan: synthetic metrics need to be checked in real scenarios, and in parallel code, a separate check for races should be built in. It makes sense to follow the fate of the fix: if Kiryl Shutsemau brings the work to a merge, it will be the first verified example of an agent-assisted perf patch in the Linux kernel.

What is still unknown / limitations

The evidence base is modest: one maintainer's email and one unmerged case — from this, one cannot draw a general capability picture of agentic coding or 'legitimization' of AI across the industry. The email is an opinion in a thread, not a documentation norm and not a patch acceptance rule; its influence on future kernel and other project guidelines remains a hypothesis. The claimed multiple throughput increase remains unvalidated, as it was obtained in a synthetic benchmark without methodology. The status of the noref-filemap-read-with-llm branch is undetermined: the race condition still needs to be fixed, and the deadline and result of Kiryl Shutsemau's work are unknown. Finally, it is unclear how the dispute over the Assisted-by trailer will end — the removal of the marking is not guaranteed, and it is precisely this that determines whether data on the provenance of AI contributions in the kernel will be preserved.

Sources

Author

Look at AI, editorial team