Debian Puts Generative AI Risk on Volunteer Submitters

Debian’s 425 unique voters chose “Responsible Use of Generative AI” on 29 August, a statement that neither endorses nor bans chatbot help in packaging, code, or docs. Each submitter still owns quality, licensing, and any leak of keys or embargoed bugs to a cloud model.

Disclosure of that help is a courtesy, not a rule, which is the part of the vote other projects did not copy.

The 55 Votes That Dropped Mandatory Disclosure

The winning text, Proposal E from Debian developer Marc Haber, is a project statement under Constitution section 4.1(5), not a new law inside the archive. It sat on a ballot with eight competing AI policy drafts plus a none-of-the-above line, after a discussion window from 23 July to 13 August and a vote from 15 to 28 August.

Haber’s option beat Lucas Nussbaum’s “Allow AI-Assisted Contributions with conditions” by 203 to 148, a 55-vote gap that decided the live rule. Nussbaum’s draft would have required contributors to flag a tool when a significant portion of a patch, mail, or bug comment was generated or substantially assisted, and it named Git trailers such as Assisted-By: as one way to do it.

Option 5 only “encourages” that flag. The same human still has to understand, review, test, and, where needed, rewrite the output. Blind uploads are called inconsistent with Debian practice. Cloud tools are told to stay away from private mail, embargoed security bugs, cryptographic keys, and credentials unless someone has explicitly authorized the disclosure.

WHAT OPTION 5 STILL DEMANDS

• Same bar: Quality, correctness, maintainability, and legal compliance do not change because a model wrote the first draft.
• Human review: The submitter is expected to understand, review, test, and modify AI-assisted output before it enters Debian.
• Secrets stay off the wire: Keys, credentials, embargoed bug detail, and private Debian mail are not to be sent to third-party AI services without authorization.
• No special AI law: Generative AI is neither exempt from nor subject to extra rules beyond the standards already expected of contributors.

Nussbaum still seconded Haber’s text, as did Tobias Frost, who wrote the more cautious Option 6 and suggested the “Responsible Use” name. The coalition on the seconds list was broader than the coalition on disclosure.

Every Permit Option Beat Every Ban

Debian Secretary Kurt Roeckx’s counting scripts finished at 00:01 UTC on 29 August. Option 5 was the only member of the Schwartz set, so it was the Condorcet winner against every rival, including none of the above.

The Social Contract ban (Option 1, from Matthias Geiger) needed a 3:1 supermajority and posted 144 against 257, a 0.560 ratio. Ian Jackson’s plan to treat LLM use as a Code of Conduct problem, including in list mail and Planet Debian posts, lost a simple majority to none of the above, 176 to 230. Those two were the only options the default beat.

OPTION 5 AGAINST EACH RIVAL

The pairwise Condorcet tally is the record those margins come from. Every option that still allowed AI-assisted work into Debian sat in the top half of the ranking; every hard restrict sat in the bottom half.

THE ROLL AND THE INBOX

• Developers on the roll: 1,045 people were eligible, with a quorum of about 48.5 votes.
• Ballots received: 575 messages hit the vote mailbox, and 137 rejects went out.
• Votes that counted: 499 passed the signature and LDAP checks, 61 were bad ballots, and 438 were tallied from 425 unique voters.

About two in five developers voted. That is a large sample by Debian GR standards, and it is still a minority of the roll. A day after the tally, the public argument was less about those cells than about the headlines that flattened Option 5 into a yes-on-AI press line. The adopted sentence is narrower: the project neither endorses nor prohibits the tools.

Who Holds the Copyright Problem Now?

Debian will not, through this vote, decide whether model output is copyrightable, or whether it is a remix of the training set. Proposal E says those questions are live in many jurisdictions, and the General Resolution does not adopt a position on them. Existing archive rules on licensing, copyright, and software freedom still apply, whatever tool produced the bytes.

The Project does not seek to resolve these unsettled legal questions through this General Resolution, nor does it adopt a position on whether AI-generated output is, in whole or in part, copyrightable or derived from copyrighted works. Instead, Debian continues to rely on the judgment and responsibility of its individual contributors.

Proposal E, Choice 5, Debian General Resolution on LLM usage

That sentence is the whole second-order move. Geiger’s failed ban argued that the Debian Social Contract and Debian Policy already demand “absolute clarity” on license and copyright, and that model output should not get a special pass. Option 5 leaves that clarity test with the person who hits send. If a contribution cannot be justified, it is not supposed to be uploaded. If it is uploaded anyway, the project has a name to hold, not a vendor and not a model card.

Geiger’s text, written with Jesse Rhodes “organically and without language model assistance,” also warned that new contributors who paste model output will burn reviewers and will not learn packaging. Option 5 answers that fear with the same review duty Debian already claims, plus a note that large automated dumps still need prior discussion. It does not add a reviewer-facing tag that would tell a tired maintainer the patch began in a chatbot.

Linux Kernel Patches Carry an Assisted-by Tag

The Linux kernel reached a nearby conclusion with a sharper paper trail. Its process docs tell humans to review AI output, keep license compatibility with GPL-2.0-only, and never let an agent add a Signed-off-by line, because only a person can certify the Developer Certificate of Origin. They also want an Assisted-by trailer on the commit.

Those kernel Assisted-by tag rules are how a maintainer knows to look twice. Fedora’s council, in a policy LWN reported as approved in October 2025, went a similar way: disclose when a significant part of a contribution is taken from a tool without changes, and treat unverified machine output as an unfair review burden. Debian’s winner cites the same human-on-the-hook idea and then declines the tag.

HOW THREE PROJECTS SPLIT ON THE TAG

Call that a family resemblance and you miss the operational gap. A kernel reviewer can grep for Assisted-by. A Debian reviewer, under the winning text, cannot. Courtesy disclosure is described as a way for people who do not want to read AI-assisted mail to choose, and as something that must not become a second quality bar. Without the tag, that choice is guesswork.

Inside a Package, the Maintainer Still Decides

Option 5 is not a command to use chatbots, and it is not a command to accept them. The statement says nothing in it requires anyone to pick up the tools, and nothing in it stops a contributor or a maintainer from refusing AI-assisted work. Debian’s usual collaboration paths, not a new AI court, are named as the way to settle those fights.

That clause is the safety valve the project-level non-ban needs. A package that wants a human-only tree can still say so. A translation team that does not want model prose can still say so. The cost of that freedom is the missing label: a maintainer who wants to refuse AI help may have to detect it from style, speed, or an admission that never arrives.

The other load the vote moved onto the same people is operational, not philosophical. Mass bug filing, large patch storms, and other automated swings across many packages still need prior talk and a human who owns the bot. The GR did not invent that expectation. It said generative AI does not relax it. For a volunteer who already triages a thousand packages, unlabeled model output plus a green light at project level is extra queue, not extra help.

Haber Used a Chatbot, Then Asked for Two Years

On 26 July, Haber offered a “deliberately permissive” draft and told debian-vote it was refined using generative AI, then edited again by hand. He also said he would not discuss the legal majority question with a model. Frost seconded it the same morning and renamed it. Eleven developers signed on, including former DPL-era figure Bdale Garbee.

After the count, Haber wrote that he regretted the loss of contributors, while noting that either winner would have cost people, and that it was too early for a final call. “Let’s revisit this in two years or so, and let this mess of a GR be a warning for our future selves,” he said on the list, in remarks carried by Slashdot. Another note on the same thread, quoted there without a name, said the winner does not acknowledge environmental cost, copyright fog, community health, or scraper load on the public web, and that a later discussion should split local models from cloud APIs.

Let’s revisit this in two years or so, and let this mess of a GR be a warning for our future selves.

Marc Haber, Debian developer and Proposal E author, debian-vote mailing list

HOW THE GR RAN

1. July 23, 2026: Discussion opens on a general resolution over LLM usage in Debian.
2. August 13, 2026: The extended discussion period closes.
3. August 15, 2026: Ranking ballots open for Debian Developers.
4. August 28, 2026: Voting closes at 23:59:59 UTC.
5. August 29, 2026: The secretary’s scripts name Option 5 as the winner.

The adopted statement already says this position describes the project at the time it is passed and that it may evolve later without another GR, while the GR process remains available if consensus fails. Haber’s two-year request and the text’s built-in drift clause point at the same unfinished argument. The vote closed. The queue of unlabeled patches, and the people who will or will not review them, did not.

Frequently Asked Questions

Did Debian change its Social Contract over AI?

No. Option 1 would have added a sixth Social Contract clause barring direct LLM-assisted contributions, which is why it needed a 3:1 supermajority and why 144 to 257 was not enough. Option 5 is only a statement under Constitution section 4.1(5), so a later shift can happen without rewriting the contract.

Who was allowed to vote on Debian’s AI policy?

Only Debian Developers on the project roll, which stood at 1,045 people when quorum was computed. Package maintainers, translators, and other contributors who are not DDs had no ballot, even though the winning text binds the work they submit.

Does the winning text treat local models differently from cloud APIs?

It does not. Option 5 talks about generative AI tools as a class and then carves out third-party services only for secrets, keys, and embargoed bugs. A follow-up on debian-vote argued that the next debate should split locally run models, where training data is easier to inspect, from cloud LLMs.

Can Debian change this AI stance without another general resolution?

Yes. The first paragraph of Option 5 says the statement describes the project at the time it is adopted and that the position may evolve as time passes without a new GR, with the GR process still available if the project cannot reach consensus. That is how Haber can ask for a two-year look-back without the winner being locked into the Social Contract.

What does the policy say about mass AI bug filings?

Large automated actions still need prior discussion through the usual project channels, the same family of process as mass bug filing in the Developers Reference. A human has to oversee the run and remain accountable for what the bot files, which means a model does not get to spray patches across the archive on its own.

Could Debian have treated chatbot use as a Code of Conduct offense?

That was Option 3, which would have required human-only drafting of bug mail, list posts, Salsa discussion, and Planet Debian blogs, with CoC discipline for violations. None of the above beat it 230 to 176, so the project declined to make LLM assistance in ordinary Debian conversation a conduct charge.