Ballots opened at midnight UTC on Saturday and close on 28 August. Eight competing proposals are on the paper, and they agree on almost nothing.
Debian developers began voting today on the most divisive question the project has taken to a ballot in years: whether code, documentation and translations written with the help of a large language model belong in the distribution at all. The Debian LLM vote runs from 00:00 UTC on 15 August to 23:59 UTC on 28 August, and there are eight separate proposals to rank.
Eight. Most general resolutions land with two or three options and a further-discussion escape hatch. This one drew a full slate after a discussion period that opened on 23 July, was extended, and finally closed on 13 August.
From a hard ban to a shrug #
Choice 1 is the proposal that started everything. Put forward by Matthias Geiger and seconded by eight developers including Ian Jackson, it would write the prohibition straight into Debian’s Social Contract, forbidding “any contributions to Debian written with the use or assistance of large language models” or other generative AI tools. The scope covers source packages, project software such as lintian, Debian web resources, documentation and translations added by Debian contributors, and official project communication.
What it deliberately leaves alone is just as interesting: upstream projects that use LLMs, AI software itself, and upstream patches or security fixes. A maintainer could still package a project whose codebase is stuffed with machine-written code. They just could not use a model to write the packaging around it. The rationale leans on reputation rather than licensing, arguing that LLM usage grows out of a “move fast, and break things” attitude that is “contrary to what makes Debian Debian”.
Choice 2 goes the other way. Rather than amend a foundation document, it issues a project statement under section 4.1(5) of the constitution allowing AI-assisted contributions subject to conditions on tooling licensing, attribution, accountability and disclosure, and it explicitly builds in room for the position to shift later “without the need to resort to future general resolutions”. It does not pretend the concerns are imaginary either. It lists them, from maintainability and legal status through environmental impact and the behaviour of AI scrapers, then argues for guardrails instead of a wall.
The middle is crowded. Choice 3 rejects LLMs “as far as practical” and moves the rule into the Code of Conduct. Choice 4 accepts AI contributions for Debian-specific work. Choices 5 and 6 are the responsible-use and cautious-approach variants. Choice 7, titled “Debian is created by humans”, tries to require a human author without dictating which tools that human reaches for. Choice 8 skips the code-quality argument altogether and objects on environmental grounds, its text careful to note that the technology itself is not the issue since it is “just maths and enough training material”, before turning on the companies pushing it.
Not every option carries the same weight #
Procedurally the ballot is lopsided. Amending the Social Contract is a foundation-document change, and Debian’s constitution sets a 3:1 supermajority for those. The statement-style options need only a simple majority. Developers rank all the choices on a Condorcet ballot rather than picking one, a system that historically rewards the option fewest people object to over the one a passionate minority loves.
The outcome matters well beyond Debian’s own archive. Ubuntu, Linux Mint, Raspberry Pi OS and Proxmox all sit on Debian, so a packaging-level rule here propagates into an enormous amount of shipped software. It also lands in sharp contrast with the Linux kernel, where AI-assisted patches are already arriving and no equivalent project-wide policy exists. Come 28 August, Debian will be the first major distribution to put an answer in writing.
Sources: the official Debian vote page and Phoronix.