The Linux kernel now has formal written rules for developers who use AI coding assistants, and the core message is that a person, not a model, answers for every patch. The guidance lives in the kernel’s AI Coding Assistants documentation, which tells AI tools to follow the same development process, coding style and patch-submission rules as any other contributor.
What the rules require
The most direct line in the document reads: “AI agents MUST NOT add Signed-off-by tags. Only humans can legally certify the Developer Certificate of Origin (DCO).” The human submitter must make sure the code complies with licensing requirements, add their own Signed-off-by tag and take full responsibility for the contribution. All code must be compatible with GPL-2.0-only.
For transparency, contributions should carry an Assisted-by tag naming the AI agent, the model version and any specialized analysis tools, such as coccinelle, sparse, smatch or clang-tidy. Basic tools like git, gcc, make and editors don’t need to be listed. ZDNet gave the example “Assisted-by: Claude:claude-3-opus coccinelle sparse” and described the tag as both a transparency mechanism and a flag that lets maintainers give AI-assisted patches extra scrutiny.
How the policy came together
According to ZDNet, the tag grew out of controversy after Nvidia engineer and longtime kernel developer Sasha Levin submitted an AI-generated patch to Linux 6.15 without telling reviewers an AI had written it. Levin later began pushing for formal transparency rules, and in July 2025 he proposed the first draft of what became the policy, initially suggesting a Co-developed-by tag. Phoronix covered that proposal, which included configuration files for popular AI coding tools that all pointed back to the same central guidance.
Good at finding problems, weaker at fixing them
A recent patch series from Arm engineer Lorenzo Stoakes shows why the human-responsibility rule matters. Stoakes used an LLM to locate single-threaded bottlenecks that slowed kernel builds, and then to work out improvements. “It generated a lot of code, much of it hideous,” he wrote, as XDA Developers reported. “I extensively audited and rewrote a lot of it, and heavily edited commit messages, the cover letter and comments.”
The finished work made allmodconfig builds up to 36% faster, incremental builds up to about 70% faster and noop builds up to about 90% faster, by Stoakes’ account. XDA noted that Linus Torvalds has recently fixed a bug with AI help, and that Debian maintainers voted to allow LLM-assisted submissions as long as the submitter takes full responsibility.
The kernel’s approach has started to travel outside the project. MakeUseOf writer Gregory Gibson applied the kernel’s tool-generated content guidance to an app he built with OpenAI Codex, recording which tool wrote which parts and keeping the specs behind the generated code. That review showed the app, which he thought worked well, was decoding audio incorrectly without raising any errors.
WebProNews is an iEntry Publication