Debian Just Put Your AI Patch on a 9-Option Ballot

TL;DR

Debian opened voting at 00:00:00 UTC on August 15, 2026 on a General Resolution titled LLM usage in Debian. The ballot carries nine ranked choices: an outright ban on LLM-assisted contributions written into the Social Contract, five varieties of "allowed, with strings", a climate-grounded objection, a humans-only rule, and None of the above. Polls close 23:59:59 UTC on August 28. The ban is the only option needing a 3:1 supermajority. Everything else needs a simple majority, which means the odds of Debian ending up with some official position on your AI patch are considerably better than even.

What is actually on the ballot

Eight proposals made it through the sponsorship process, plus the mandatory None of the above. The discussion period ran from July 23 to August 13 and was extended once, which for a Debian governance thread is practically a rounding error.

Here is what each proposal wants, and how many Debian Developers seconded it onto the ballot:

Option H, from Holger Levsen, argues that "global LLM usage accelerates the destruction of our ecosystem (planet earth) and that is a deal-breaker". It collected seventeen seconds, more than double the outright ban, and it is the only option on the ballot that spends its rationale on power draw rather than on patch review.

Do not read those bars as a poll. Seconding a Debian proposal means "this belongs on the ballot", not "I will vote for it". Pierre-Elliott Bécue seconded the outright ban and then proposed option D, which explicitly permits AI contributions for Debian-specific work. Ian Jackson seconded the ban too, then filed option C as a softer alternative. The seconds count is a map of what the project wanted argued in public, not a forecast.

Option A is the one with teeth

Matthias Geiger's proposal is the only one that changes a foundational document. It amends the Social Contract to "expressly forbid any contributions to Debian written with the use or assistance of large language models (LLMs) or other generative AI tools". The rationale is blunt about why: LLM usage, it argues, comes from a "move fast, and break things" attitude that is "contrary to what makes Debian Debian".

The scope is narrower than the headline suggests, and the boundary is the interesting part.

That carve-out matters more than it looks. Debian packages tens of thousands of upstream projects, and a growing share of them accept AI-assisted patches. Even the strictest option on the ballot only governs what Debian Developers themselves write. It is a ban on the shipping label, not on the contents of the box.

Why 3:1 is a real wall

Debian counts with a ranked Condorcet ballot, resolved through the Schwartz set as defined in the project constitution. Normally an option just has to beat the default in a pairwise comparison. A supermajority requirement changes the arithmetic: per section A.5.3, an option needing S:1 must win its pairwise comparison against the default by a factor of S, so option A has to be preferred to None of the above by at least three votes to one.

Think of it as a photo finish where one horse has to win by three lengths instead of a nose. Constitutional amendments in Debian are supposed to be hard, and this is what "hard" looks like in practice.

GCC already ran this play, three weeks ago

Debian is not deciding in a vacuum. On July 29, 2026, five days into Debian's discussion period, the GCC steering committee adopted its own AI policy: GCC declines "legally significant contributions which include LLM-generated content or are derived from LLM-generated content", using the GNU maintainer guidelines threshold of roughly 15 lines of code or text. Test cases get an explicit exception, and using an LLM for research, analysis, bug hunting or patch review stays fine as long as its output does not land in the tree.

GCC's reasoning is licensing, not taste. If AI-generated code has no human author, it has no copyright holder, and a project that enforces the GPL by asserting copyright over its own tree does not want a hole in that chain. Several of Debian's options reach for the same argument. Option B requires contributors to verify licensing compatibility and take full responsibility for what they submit; option E puts the responsibility on the contributor without demanding disclosure at all.

Why this matters if you have never filed a Debian bug

Debian sits upstream of Ubuntu and, by extension, of a very large share of the container images your agents are running inside right now. A project-wide stance from Debian is the closest thing the Linux ecosystem has to a reference implementation of AI contribution policy, and the smaller projects downstream tend to copy rather than re-derive.

The practical questions this vote answers, one way or another:

• Disclosure. Options B and D want AI assistance flagged in commit metadata. Option E explicitly does not require disclosure. Whichever wins sets the default expectation for how you label agent-written patches.
• Scope of "contribution". Option G targets the AI output itself while leaving contributors free to use assistive tools, a distinction that will get tested the first time somebody pastes a generated commit message.
• Who eats the review cost. The subtext across every proposal is volunteer reviewer time. A patch that took eleven seconds to generate can take an hour to review, and Debian's reviewers are unpaid.

Caveats worth keeping straight

Nothing is decided. Voting is open until August 28 and Debian ballots are frequently won by None of the above or by an option nobody predicted. The electorate is small: the 2026 DPL election drew 347 votes, so a few dozen developers changing their ranking can move the result.

The Project Secretary also issued a corrected ballot at 02:14 UTC on August 15, roughly two hours after polls opened, which is a good reminder that even a project with a constitution, a Condorcet tally and a thirty-year paper trail still ships a hotfix now and then.

Key Takeaways

• Debian is voting from August 15 to August 28, 2026 UTC on nine ranked options covering LLM use in the project.
• Only option A, an outright ban written into the Social Contract, needs a 3:1 supermajority. The other seven need a simple majority.
• Even the ban would cover only Debian's own packaging, project software, docs and communications. Upstream projects using LLMs are explicitly out of scope.
• The climate-based objection drew 17 seconds, the most of any option, though seconds indicate ballot access rather than support.
• GCC set a precedent on July 29: no legally significant LLM-derived contributions, using the GNU guidelines threshold of about 15 lines, with a test-case exception.
• Debian's answer will propagate downstream to Ubuntu and to the base images most AI infrastructure runs on, so the outcome is worth watching even if you never touch a .deb.

Sources: Debian General Resolution: LLM usage in Debian, debian-vote first call for votes (corrected ballot), Debian Constitution, Phoronix, LWN on the GCC AI policy, Debian DPL election 2026 results