Nixpkgs contributors push to ban LLM text outside code

A NixOS contributor has called for an explicit ban on large language model output in pull request descriptions, commit messages, review replies, and other non-code text in nixpkgs, after two PRs arrived with multi-hundred-word walls of prose.

Ethan Edwards pointed to a pair of submissions whose descriptions ran to roughly 800 and 660 words, with commit messages totaling hundreds more. He argued the author likely had not fully read or understood the changes, and that any review would simply be routed back through a model. Even if such work meets the letter of the project's existing AI and automation policy, Edwards said it violates the spirit by shifting the entire burden of understanding onto reviewers: extractive contributions that take the human out of the loop.

The proposal would still allow LLM use in code itself but bar generated text everywhere else, on the grounds that contributors must show they grasp what they submit. Responses on the NixOS Discourse split along familiar lines. Some backed a hard ban on LLM-assisted contributions of any kind. Others said the real problem is length and structure, not provenance, and favored enforceable word limits, strict PR templates, or bots that close oversized or non-compliant submissions before a maintainer opens them.

Detection is the sticking point. Lightly edited model prose is hard to distinguish from human writing, so a ban on "LLM-looking" content would rest on weak signals. Autistic and neurodivergent contributors already risk misclassification by detectors and by readers who associate certain styles with generated text, which several people called an unfair barrier. A common counter was that unstructured walls of text are unwelcome regardless of origin, and that reviews can simply demand shorter, clearer writing.

The project's steering committee may put the broader AI policy to a referendum. Until then, nixpkgs still has to decide whether to tighten rules on generated prose, cap verbosity, or keep leaving extractive work for individual reviewers to reject.