This is Part 7 of AI Grew Up on Git, a series on the Git infrastructure that raised AI and what AI is now doing to it.
In June 2025, at an Open Source Summit, an NVIDIA engineer named Sasha Levin told an audience something that had already happened without anyone noticing. A patch of his, already merged into the Linux kernel by a maintainer who trusted his name, had been written with the help of a large language model. No disclosure had gone with it. Nobody had asked. It sat in the tree like every other patch, indistinguishable, until Levin said so out loud.
The reaction was immediate and it was not gentle. A merged patch in the world’s most scrutinized codebase had an origin nobody had been told about, and the discomfort wasn’t really about that one patch. It was about the fact that nothing had stopped it, and nothing would stop the next one either, because the entire review process assumed a fact that had quietly stopped being reliable: that the person whose name was on a change was the one who had written it.
The patch nobody disclosed
The story could have ended there, another scandal that produces outrage and then nothing. Instead the person who caused the problem became the person who fixed it.
In July 2025, Levin posted a proposal to the kernel mailing list, an early, technical idea: a unified way for AI coding tools to find the kernel’s contribution rules, symlinks pointing every AI assistant’s config file at one shared document. It was a small, practical fix for a much larger question the kernel hadn’t answered yet, and posting it in public is how kernel policy always starts, as an argument other people get to have.
Two years on a mailing list
The argument ran for the better part of two years, in the open, the way this project settles everything.
Intel’s Dave Hansen pushed back with a revised draft that emphasized transparency over convenience. Lorenzo Stoakes argued a scattered set of config files wasn’t good enough, that the project needed an actual policy document. Michal Babka warned that shipping the configs alone, without a document behind them, might quietly tell developers they didn’t need to think about any of this. The 2025 Linux Kernel Maintainers Summit finally converged on three points: a human is accountable for every patch, no matter what wrote the first draft; purely machine-generated submissions are not welcome; and AI assistance must be disclosed.
On December 23, 2025, Sasha Levin submitted the patch that turned those three points into fifty-nine lines of documentation, Documentation/process/coding-assistants.rst, its commit message citing the Summit’s consensus by name. Jonathan Corbet, the kernel’s documentation maintainer, replied the same day. His answer was short: it looked good to him, and if nobody objected, he’d merge it before long. Nobody objected. It shipped as part of Linux 7.0, released by Torvalds that April. Two years of argument, closed by one conditional sentence and silence.
