The Linux kernel project has dropped its requirement to label LLM-assisted commits with the 'Assisted-by' tag, a reversal from the policy introduced in April 2026. Jakub Kicinski confirmed in an August LKML post that budget has been secured to run Meta's LLMs directly on patch review, commit message creation, and editing. Multiple frontier models are being run against each patch simultaneously to filter out confabulations.
A new bot called 'Sashiko' has also been introduced to handle code reviews. Linus Torvalds defended the direction in a July LKML post. The decision is not unanimous among maintainers, but Phoronix reports it is proceeding regardless. The argument inside those mailing list threads, including how confabulation filtering actually works at scale, is worth reading in full.
Most open source projects have moved the opposite direction, banning or tightly restricting LLM contributions on grounds of slop, copyright, and licensing risk. The Linux kernel is now the highest-profile counterexample. The outcome will be a real-world test of whether frontier model tooling can hold up inside the most scrutinized codebase in existence.
[READ ORIGINAL →]