Debian’s general resolution on generative AI closed August 28, 2026, with a clear winner: choice 5, “Responsible Use of Generative AI,” whose first substantive line reads “Debian neither endorses nor prohibits the use of generative AI tools.” 425 of 1,045 eligible Debian Developers voted during the two-week ballot window (August 15 to 28). The winning option is the Condorcet winner of a nine-option ballot - it beat every rival head-to-head, including the closest one by 55 votes.
Nine options, one winner #
The general resolution drew eight substantive proposals plus “None of the above,” spanning the full range from total ban to unconditional acceptance:
- Choice 1 (Matthias Geiger): a Social Contract amendment that would ban LLM-written contributions outright.
- Choice 2 (Lucas Nussbaum): AI-assisted contributions allowed under conditions, including a
`Generated-By:`
git trailer, prior discussion before bulk changes, and no sensitive data sent to untrusted providers. - Choice 3 (Ian Jackson): reject LLMs as far as practical, mandatory disclosure, violations treated as Code of Conduct breaches.
-
Choice 4 (Pierre-Elliott Becue): accept AI contributions for Debian-specific work only, with signed submissions and marking.
-
Choice 5 (Marc Haber): neither endorse nor prohibit; existing contributor standards apply (the winner).
-
Choice 6 (Tobias Frost): a cautious approach - avoid where practical, disclosure as a courtesy rather than an obligation.
-
Choice 7 (Gard Spreemann): disallow generative AI “output” as direct contributions, allow assistive use.
-
Choice 8 (Holger Levsen): an anti-LLM position statement on climate grounds - “Fight the game, but not the players.”
What the winning statement says #
The statement is a position of the project under section 4.1(5) of the Debian Constitution, so it can evolve without a new vote. Its substance:
- Every contribution must meet the same standards of quality, correctness, maintainability, and legal compliance as any other, regardless of the tools used to produce it.
- “Blindly accepting or up AI-generated material” without human review is called out as inconsistent with Debian’s practices.
- Disclosure of AI assistance is encouraged but not required. Nussbaum’s conditional option, which lost the final pairwise contest, had made a
Generated-By:
trailer the mechanism; nothing in the winning text mandates one. - Contributors must not send confidential or security-embargoed material to third-party AI services.
- Mass bug filing, wide refactors, and other large-scale automated changes need prior discussion on the lists and a human overseeing them.
- The statement explicitly declines to resolve the unsettled legal questions around AI-generated material’s copyright and licensing, leaving those judgments to contributors.
The net effect: no special rules beyond the standards already expected of Debian contributors. One observer on X put the caveat precisely: saying the new Debian policy is in favor of AI use is a bit misleading - the text neither endorses nor prohibits. The kernel’s own stance, which permits AI-assisted patches when they are tagged and reviewed, is the closest precedent.
joey, a Debian Developer, then pointed at the text itself: the winning proposal was drafted with LLM assistance. Ted Ts’o walked through how that passed the smell test: a human conceived the ideas, an LLM drafted the wording, and a human edited and took responsibility for the result - “a human took responsibility for the final result.”
The tally #
The two sharpest restrictions were the only two options that failed to beat “None of the above.” The ban (choice 1) needed a 3:1 supermajority and scored 0.560. Jackson’s choice 3 scored 0.765 against the required simple majority of 1.0. Every other option cleared its majority check, and choice 5 then beat them all in pairwise contests. Commenter bluca read the defeat of choices 1 and 3 as the electorate speaking in one register: “no, absolutely not, go away.” Commenters mhvk and hlieberman both read the result as a status-quo win rather than a mandate: roughly 30 percent of voters preferred a no-AI option.
The resignation #
On August 29, one day after the result, Debian Developer Antoine Le Gonidec retired from the project. His resignation, first sent to the private developers list and then shared publicly, is direct about the reason: “I can not support the current decision of Debian about LLM use, and am no longer willing to be seen as a part of Debian under these new rules.”
His words, not ours: “I want the consequences of Debian fully embracing fascist-backed technology to be known.” And: “Pretending to have a ‘neutral stance’ when faced with fascism is not neutrality, it’s active collaboration.” He signed off with “You can replace me with some agentic AI anyway, so do not even try to pretend you’re going to miss me,” and closed the door by asking not to be CC’d on any replies.
The split #
Brodie Robertson’s read on the divide: “There is clearly a very large cultural divide within the project that a policy change doesn’t resolve.” He followed with the historical precedent - Devuan exists because the systemd debate split Debian’s developers in 2014, and a fork over LLM usage is not physically impossible. The counter-argument, from lawyer Greg Helding on the same thread: “Voting against AI completely is nonsensical. AI is used in kernel development. A distro that wants to be AI-free would have to fork the kernel.” A true AI-free distro cannot even ship Linux without patching out upstream’s accepted practice.
The resignation is not the whole register of dissent. Commenter dskoll was sad but called the outcome “inevitable,” because Debian is “too large a project to be able to meaningfully enforce a stronger stance against LLMs.” Ted Ts’o reached for the dot-com bubble: bubble economics around a technology do not make the technology itself evil, and radical rejection of AI would repeat the mistake of rejecting the internet over e-commerce’s excesses. And hmh’s objection is the one that touches agent tooling directly: the resolution protects humans using LLMs but contains nothing that stops “fire-and-forget” AI agents from spamming maintainers. An anti-agent-harassment provision was on nobody’s ballot.
Omarchy’s $12M month #
The vote landed the same month DHH’s Omarchy project - an Arch- and Hyprland-based distro built around agent-first tooling - turned itself into a funded non-profit. The Omacom Foundation launched August 21, 2026 with $8M committed, and grew to $12M from twelve Founding Patrons by August 31, each putting in $1M of personal money: Tobi Lutke, Patrick Collison, Michael Dell, Jack Dorsey, Matthew Prince, Brendan Iribe, Jason Fried, Drew Houston, Peter Steinberger, Brian Armstrong, Yunjie Dai, and DHH. The foundation holds the trademarks, funds infrastructure, and sponsors the open-source work Omarchy depends on; its first named sponsorship went to Hyprland creator Vaxry.
One more thread worth noting for where this goes next: when Maks Surguy suggested the foundation “should get AI token donations as well as cash,” DHH replied “Announcement coming soon on that.”
How other projects landed #
The spectrum around Debian: the Linux kernel allows AI-assisted patches with disclosure tags, and is simultaneously absorbing the cost of AI scrapers. Kernel.org infrastructure maintainer Konstantin Ryabitsev reported this week that legitimate traffic is down to about 2 percent of requests to git.kernel.org. Anubis proof-of-work checks now bat away 66 percent of challenge attempts, but a third of bots solve the math; after the difficulty was raised to 5, bots started solving difficulty 5, and the blocking failure moved to residential proxies rented out through smart-TV app SDKs. Meanwhile 14 to 16 of the fleet’s 90 CPU cores - five geo-distributed nodes - constantly render commits as HTML for scraper bots, roughly 20 percent of total capacity, most of it on commit and log pages.
GCC bans AI-generated code from the compiler itself. Oracle’s OpenJDK ban on AI-generated contributions stands. Debian now sits at the kernel end of that spectrum rather than the GCC end.
What it means for local AI developers #
One rule has teeth: if you handle embargoed security material, keep it away from hosted AI services entirely. That is the complete list of new obligations.
A local model clears that rule automatically, because the material never leaves the machine. Everything else works the way it did before the vote: no disclosure trailer is required, nobody checks which tools wrote a given patch, and reviewers grade the patch itself, not its authorship. The statement does not tell you to run a hosted model or a local one - that choice only matters for the confidentiality rule, and local hardware satisfies it by default.