Debian Is Voting on Whether AI May Touch Its Code
One of the oldest Linux projects has eight options on the ballot, from a full ban on LLM-assisted contributions to a permissive framework. Voting runs until 28 August.
Debian, the volunteer Linux distribution that quietly sits underneath Ubuntu and a large share of the world’s servers, opened a formal vote on 15 August about how large language models may be used inside the project. The ballot stays open until 28 August, and only Debian Developers, the project’s accredited members, can cast a vote.
There are eight proposals plus a “none of the above” option, and they span the full range. Proposal A, put forward by Matthias Geiger, is the strictest: a ban on LLM-assisted source code, web resources and official project communications. Proposal B, from Lucas Nussbaum, is the most permissive and would allow AI-assisted work with guardrails such as disclosure and human review. An eighth proposal comes at it from a different angle entirely and focuses on the energy cost of running these models, asking contributors to avoid them where they reasonably can. Debian uses ranked voting, so members order the options by preference rather than picking one.
A detail worth noting: at least one of the restrictive proposals carves out an exemption for upstream projects, meaning code Debian packages but does not write. Without that, the policy would collide with the Linux kernel itself, where a meaningful share of new submissions is already AI-assisted.
Here is why this is bigger than one distribution’s internal paperwork. Open-source projects run on trust and on the idea that every line has a human who understands it and can be asked about it. AI-generated patches strain both. They can look plausible and be subtly wrong, they arrive faster than reviewers can read them, and the licensing of the training data behind them is still legally unsettled. Debian is unusually deliberate about governance, so whatever it decides will become the reference document other projects argue over. The discussion phase alone ran from 24 July and produced five competing drafts before the ballot was finalised, which tells you how genuinely split the community is.
What this means for you: if you do not contribute to open source, nothing changes on your machine. If you do, this is the moment to read your own projects’ rules, because plenty have none and are about to need them. The practical middle ground most maintainers are converging on is simple and worth copying: use whatever tools you like, but disclose that you did, and never submit code you could not defend line by line in review. And if you follow AI policy debates generally, keep an eye on the tally after 28 August. This is one of the first times a large, old, genuinely democratic technical community has put the question to a formal vote instead of leaving it to a founder’s opinion.
Sources
Google Open-Sourced the Code That Checks Where an Image Came From
Credentio is a C++ library for verifying C2PA Content Credentials on your own device, with nothing sent to a server. The same code already runs inside nearly 40 Google products.
