vikrant69g blog

Debian might ban AI-generated code contributions

A Debian general resolution proposes banning LLM-generated patches. The reasoning is direct: you cannot verify the training data's licensing.

Debian swirl logo on a dark background with binary code overlay

Debian is voting on a general resolution to ban code contributions generated by large language models. Not discourage. Ban. The argument is simple. Every line of code in Debian must have a clear licence. If you submit a patch, you are asserting you wrote it or you know where it came from. LLMs do not work that way. The model ingests millions of files, some GPL, some MIT, some proprietary, some scraped without permission. When it generates a function, you cannot prove it is not a memorised chunk of someone else’s code. This is not theoretical concern. GitHub Copilot has been shown to reproduce exact snippets from training data, complete with comments and quirks. If that snippet is GPL-licensed and you paste it into a BSD project, you just contaminated the licence tree. Debian’s position is that they will not accept that risk. The resolution does not ban using LLMs as a learning tool or for drafting ideas. It bans submitting the output as a patch unless you can verify every line’s provenance. That is a standard no one can meet with current models. This puts Debian ahead of most open-source projects, which are still pretending the problem does not exist. The Linux kernel has no formal policy. Python’s CPython repository has no policy. Most projects are waiting to see if someone gets sued first. What makes this interesting is enforcement. How do you prove a patch was AI-generated if the submitter does not admit it? You cannot. The resolution is as much a statement of values as a technical control. Debian is saying the burden of proof is on the contributor, not the maintainer. If this passes, expect other distributions to copy it. The DFSG has always been stricter than most open-source definitions. That strictness is why Debian is still the base for Ubuntu, Kali, and half the server installs in production. Trust matters more than convenience.


Source: GR: Ban LLM Contributions from Debian