Debian Just Put Your AI Coding Habits on a Ballot. The Ban Option Has to Win 3 to 1.

TL;DR

On July 24, Debian opened the formal discussion period on General Resolution 2026-002, "LLM usage in Debian." Four options are on the ballot: ban LLM-assisted contributions outright, allow them under six conditions, reject them as far as practical, or accept them with mandatory disclosure. The ban option would add a new section to the Debian Social Contract, and under the project's constitution that means it does not need to win, it needs to win three to one. Debian sits upstream of Ubuntu and most of the derivative world, so this ballot will not stay on debian-vote.

What is actually on the ballot

The trigger was a message posted to debian-vote on July 22 at 17:35 CEST by Matthias Geiger, titled "GR: Ban LLM contributions from Debian." Within two days it had attracted three competing ballot options and enough seconds for all four to qualify. The discussion period formally opened July 24.

Option A, Geiger's ban. It amends the Social Contract with a new section 6, "Works Created through the use of Large Language Models (LLMs)":

We will not allow direct contributions to Debian written with the use or assistance of large language models (LLMs) or other generative AI tools. Direct contributions are defined as packaging, native Debian software like lintian, documentation and translations written by Debian contributors, and official Debian web resources, etc.

Upstream software that uses AI, and upstream patches and security fixes, are explicitly out of scope. The ban is about work done by Debian, not work Debian ships.

Option B, Lucas Nussbaum's permissive path. Allowed, subject to six conditions: the tool's terms must not conflict with Debian's right to distribute and modify, third-party licensing and attribution must be verified, the contributor holds full accountability for technical merit and security, significant assistance must be disclosed (a Git trailer such as Generated-By or Assisted-By), bulk or automated changes must be discussed in advance, and tools that transmit sensitive project data are out.

Option C, Ian Jackson's middle finger to the middle ground. Avoid LLMs where practical, disclose any use, respect individual teams' own bans, and, the part that will get quoted, prohibit LLM assistance in messages to humans: emails, bug reports, blog posts. Under option C you could arguably still generate a patch, but you would have to write the bug report about it yourself.

Option D, Pierre-Elliott Bécue's guidelines. Accept AI-assisted work for Debian-specific tasks with the submitter carrying responsibility, require DFSG compliance and genuine understanding of the change, mandate marking AI assistance in commit messages, and ban cloud-based AI for sensitive or private data.

Read the sponsor lists and the picture gets stranger. Geiger, who proposed the ban, seconded Jackson's option C. Jackson, who proposed C, seconded the ban. Bécue proposed the permissive option D and also seconded the ban. In Debian, a second is ballot access, not a marriage proposal, and developers routinely sponsor options they intend to rank below "further discussion." Nobody has voted yet. Everybody has already been counted.

Why the ban has the hardest math

Here is the detail that most coverage is skipping, and it decides the outcome. Options B, C and D are position statements. Option A edits the Social Contract, which is a foundation document, and section 4.1.5 of the constitution says a foundation document "requires a 3:1 majority for its supersession."

The difference is the difference between passing a bylaw and amending a constitution. A bylaw needs the room to mostly agree. An amendment needs the room to overwhelmingly agree, and a determined quarter of the room can stop it cold. In a project that has spent five months failing to reach consensus on this exact question, a quarter is not a high bar to clear.

The case the ban makes

Geiger's proposal lists four objections, and they are not vibes. Copyright first: LLM output has, in his words, "very unclear legal status," which collides with the DFSG requirement that Debian know exactly what license every byte in the archive carries. Then quality, with a specific and very Debian complaint: generated packaging arrives with non-functional watch files and invented copyright information. A watch file that does not work is a packaging bug. A copyright file that hallucinates an author is a legal claim.

Third, reviewer load. Generating a package became nearly free; reviewing one costs exactly what it always did, and the proposal names the outcome as burnout. This is the same arithmetic Godot used on June 30 when it banned substantial AI-authored code outright. Fourth, ethics and infrastructure: crawlers scraping for training data have been hammering Debian's own servers, which turns an abstract argument about consent into a concrete argument about bandwidth bills.

Debian already tried this once, and blinked

February 2026: Nussbaum opened a draft GR to allow AI-assisted contributions, prompted by what he described as attacks on people using AI inside the project. It went nowhere. On March 3 he concluded that as long as the discussion stayed calm, Debian could keep handling cases individually. LWN summed it up on March 10 under the headline "Debian decides not to decide on AI-generated contributions," and the project went back to case-by-case judgment under existing policy.

That truce lasted about four months. The difference this time is direction: the February draft asked permission to allow, the July proposal demands a prohibition, and prohibition is what gets people to the ballot box.

What happens next, and when

The constitution sets a minimum discussion period of two weeks (section A.1.1) and a voting period of two weeks that the Project Leader may vary by up to one week (section 4.2.3). No voting date has been set. Discussion periods get extended when amendments keep landing, and amendments are landing: option B already carries two of them.

Volume gives you a sense of the temperature. As of 07:40 UTC on July 26, the July archive of debian-vote held 189 messages, on a list whose entire purpose is procedural. Debian's electorate is real too: 568 ballots were cast in the 2026 Project Leader election. This is not a forum thread with a poll attached.

Why you should care if you do not run Debian

Because the ballot options are the first serious attempt by a large, old, deliberately slow institution to write down what "AI-assisted contribution" even means, and the vocabulary tends to get copied. Option B's Git trailers (Generated-By, Assisted-By) are a mechanism any project can adopt tomorrow. Option A's carve-out, banning LLM work done by Debian while still shipping upstream software written with it, is the line every distro and every enterprise repo will eventually have to draw somewhere.

The practical read for anyone shipping agent-generated patches into open source: disclosure is winning on the merits even where the ban is not. Three of the four options require you to say so. None of them let you quietly ship a diff you cannot explain, which was the actual complaint all along.

Key Takeaways

• Four options, live now: ban (A), allow with six conditions (B), reject as far as practical (C), accept with disclosure (D). Discussion opened July 24, 2026.
• The ban needs 3:1, not 50%: option A amends the Social Contract, a foundation document, and section 4.1.5 of the constitution requires a 3:1 majority to supersede one.
• Seconds are not votes: the ban's proposer seconded a rival option and vice versa. All four cleared sponsorship; none of that predicts the tally.
• The complaints are concrete: broken watch files, invented copyright entries, reviewer burnout, and scrapers hitting Debian's own infrastructure.
• Debian punted on this in March after a February draft fizzled, and the truce held roughly four months.
• Disclosure is the consensus even without a ban: three of the four options require marking AI assistance, typically in a commit trailer.

Sources: Debian General Resolution: LLM usage in Debian, Matthias Geiger's original proposal on debian-vote, debian-vote July 2026 archive, Debian Constitution, Debian Social Contract, LWN: Debian decides not to decide on AI-generated contributions, Phoronix, Linuxiac