The Linux kernel project has established rigorous documentation and attribution requirements for AI-assisted code, mandating human accountability for all automated contributions.
The Linux kernel community has officially codified its stance on the integration of large language models into its development workflow, establishing strict documentation requirements for AI-assisted code contributions. This formalization, finalized following extensive debate at the 2025 Maintainers Summit, delineates the boundary between automated assistance and human authorship within the kernel tree.
Sasha Levin, a Linux LTS co-maintainer at NVIDIA, initiated the process in mid-2025 by proposing a standardized set of configuration files and guidelines for AI tools. These efforts culminated in the inclusion of documentation within the kernel source at Documentation/process/coding-assistants.rst. The guidelines stipulate that AI tools must adhere to the same rigorous standards as human contributors, including full compliance with the project’s established coding style and submission protocols.
Central to the new policy is the absolute prohibition of AI agents signing off on code. The documentation explicitly states that AI agents must not add Signed-off-by tags, thereby ensuring that a human contributor retains total legal and technical responsibility for every patch. This requirement reinforces the project’s reliance on the Developer Certificate of Origin, which necessitates human verification of license compatibility and code integrity.
Transparency is enforced through the mandatory inclusion of an Assisted-by tag in commit messages. This tag must specify the AI agent, the model version, and any ancillary analysis tools utilized during the development process. Maintainers utilize this metadata to track the prevalence of AI involvement and to determine when a submission requires heightened scrutiny during the review cycle.
Greg Kroah-Hartman has already implemented these principles by utilizing AI-assisted fuzzing in his own tree to identify potential vulnerabilities in ksmbd and SMB code. While these automated tools effectively surface latent issues, the staging subsystem now rejects most LLM-generated patches to prevent the accumulation of low-quality code. This selective approach ensures that automation serves as a diagnostic aid rather than a primary source of kernel modifications.
The practical application of these tools has yielded mixed results, as noted by Arm engineer Lorenzo Stoakes. While Stoakes successfully leveraged an LLM to diagnose single-threaded bottlenecks, he emphasized that the generated code required extensive manual auditing and rewriting before it met kernel standards. His efforts ultimately resulted in significant performance gains, including a 36 percent improvement in allmodconfig build times and a 90 percent reduction in noop build times.
Linus Torvalds, who previously expressed skepticism regarding AI-generated code, has integrated these tools into his own debugging workflows. During a complex session involving Intel graphics drivers, Torvalds utilized an LLM to manage repetitive test code generation while maintaining human oversight of the final implementation. This experience highlights a recurring theme among kernel developers: AI excels at exploration and exhaustive testing but remains deficient in the nuanced judgment required for kernel-level architecture.
The influx of automated submissions has introduced significant operational challenges for maintainers, who must now triage a higher volume of patches. Networking maintainer Jakub Kicinski reported that up to half of the patches in the net-next tree during the 7.3 cycle were linked to automated or AI-driven cleanups. This surge in volume has necessitated more robust filtering mechanisms to separate high-quality contributions from noise, leading to the removal of some drivers that became targets for automated checkers.
The project’s stance reflects a broader strategy to manage the risks associated with training data opacity and potential license violations. Every contribution must remain compatible with the GPL-2.0-only license, and contributors are solely responsible for ensuring that no copyrighted material is introduced by AI models. By channeling the energy of automated tools while maintaining strict human gatekeeping, the Linux kernel project aims to preserve its long-standing standards despite the rapid proliferation of generative coding assistants.
Future development will likely focus on refining these guidelines as maintainers gather more data on the long-term impact of AI on release cycles. The community continues to monitor the trade-off between the ability of AI to surface corner cases and the increased burden placed on human reviewers to filter the resulting output.
