Last updated: August 30, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
Debian's two-week vote on AI-assisted contributions closed August 28, 2026, and Debian Developers picked "Responsible Use of Generative AI," a proposal from longtime developer Marc Haber, over seven rival options including an outright ban. The policy doesn't ban AI tools and doesn't require using them either — it just makes the human who hits "submit" fully responsible for whatever came out of the model.
Short answer: On August 28, 2026, Debian Developers voted for General Resolution 2026-002, "Responsible Use of Generative AI." The winning option beat the runner-up 203–148, then beat the rest by wider margins. Debian now neither endorses nor bans AI tools in project work; contributors must review, test, and take full responsibility for anything AI-assisted, and disclosing AI use stays optional, not required.
I covered this fight when it opened two weeks ago in my piece on Debian’s vote on AI and LLM contributions, and I bet against a full ban clearing the 3:1 supermajority bar while guessing a disclosure-and-accountability option would come out ahead. That's roughly what happened, just not the exact proposal I expected. In my testing this week, I ran the same three tasks — a packaging patch, a changelog entry, and a disclosure trailer — through three AI coding assistants to see whether any of them handle the "review, test, and take responsibility" half of Debian's new policy on their own. None do. That part is still entirely on the human, which is the whole point of how this resolution is written.
What actually won, and by how much
This was a General Resolution (GR). It's the mechanism Debian uses to settle project-wide policy that no single team can decide alone. Eight formal proposals, lettered A through H, went to a vote. Debian counts these votes Condorcet-style: developers rank every option they find acceptable instead of picking just one, and the system finds whichever option beats every other option head-to-head.
Marc Haber's Option E, "Responsible Use of Generative AI," won that comparison. According to the official results on debian.org, it beat four rivals by increasingly wide margins: Lucas Nussbaum's more conditional Option B, 203 to 148; Pierre-Elliott Bécue's narrower Option D, 232 to 115; Tobias Frost's more cautious Option F, 210 to 130; and Gard Spreemann's "Debian is created by humans" Option G, 251 to 139. The two strictest anti-AI proposals fared worse still. Matthias Geiger's Social Contract ban (A) and Ian Jackson's "reject LLMs as far as practical" (C) didn't even out-poll "None of the above" in the final tally. Roughly 64% of developers ended up preferring some form of permission over a ban or strong discouragement, per Phoronix’s count of the results. That's a real split, but not the coin-flip the mailing-list arguments made it sound like.
Nussbaum, whose own proposal came in second, summed up the mood afterward: "We have a lot of work ahead of us to act as a community and understand how we can continue to accommodate the very large minority that would have preferred to ban or discourage AI." Haber himself was even more blunt about how close and contentious the process had been: "It is just too early to take a final decision on the matter. Let's revisit this in two years or so, and let this mess of a GR be a warning for our future selves."
What the policy actually requires
The operative text is short. It's worth reading in full rather than trusting a headline. Per the resolution: "Debian neither endorses nor prohibits the use of generative AI tools in the development, maintenance, or documentation of software, packaging, documentation, and other media published within the Debian Project." Everything else in GR 2026-002 is guardrails around that one sentence.
- Existing standards still apply. A patch, doc page, or translation is judged the same way regardless of whether a human or a model drafted the first pass. AI-assisted work that fails DFSG licensing, review, or quality bars gets rejected exactly like human-written work would.
- Responsibility doesn't transfer to the tool. The text is explicit: "The use of a generative AI tool does not diminish the contributor's responsibility for the work they submit." Contributors are expected to understand, review, test, and modify AI output before it ships.
- Disclosure is encouraged, not mandated. Unlike Nussbaum's runner-up proposal, which would have required a "Generated-By:" commit trailer, the winning text leaves disclosure voluntary.
- Confidential data stays out of third-party AI services. Security embargoes, private communications, and non-public Debian material can't be pasted into an external model without authorization.
- Large-scale automated changes need a human in the loop. Any AI-driven process touching many packages at once requires prior discussion and a person who stays accountable for what it does.
- The copyright question stays open. The resolution notes the legal status of AI-generated material "remains the subject of ongoing discussion" and doesn't try to settle it unilaterally.
Where the eight proposals landed
| Option | Sponsor | Core stance | Result vs. the winner |
|---|---|---|---|
| E (winner) | Marc Haber | Neither endorse nor prohibit; contributor stays responsible | — |
| B | Lucas Nussbaum | Allow with conditions, incl. mandatory disclosure trailer | Lost 148–203 |
| D | Pierre-Elliott Bécue | Accept AI use for Debian-specific work only | Lost 115–232 |
| F | Tobias Frost | Cautious approach; discourage without banning | Lost 130–210 |
| G | Gard Spreemann | "Debian is created by humans"; AI for research only | Lost 139–251 |
| A | Matthias Geiger | Full ban via Social Contract amendment (needed 3:1) | Behind "None of the above" |
| C | Ian Jackson | Reject LLMs "as far as practical" | Behind "None of the above" |
| H | Holger Levsen | Avoid LLMs; environmental cost is a deal-breaker | Placed near the bottom |
Example prompts you can copy
If you contribute to Debian, or maintain any project that's about to have this same argument, these are the checks Haber's policy effectively expects you to run yourself, since nothing does it automatically:
- Explain-it-back check: "Walk through this patch line by line and explain what each change does and why. Flag anything I can't justify without re-reading the AI's original output."
- Voluntary disclosure line: "Draft a Debian-style changelog entry for this AI-assisted patch, and add an optional line noting AI assistance in the same tone as an existing entry, not as a formal disclaimer."
- Confidentiality check: "Before I paste this file into an AI tool, tell me if anything in it looks like a security embargo, credential, or private communication that shouldn't leave this machine."
- Licensing sanity check: "Review this AI-drafted documentation paragraph for phrasing that looks lifted from a specific external source, so I can check it against DFSG licensing before I submit it."
Common mistakes to avoid
The mistake I saw most in the days after the vote is reading this as "Debian now requires AI disclosure." It doesn't. Nussbaum's Option B would have, and it lost. Disclosure is encouraged, full stop. A maintainer can still ask a contributor to disclose on a given package, but the project-wide rule doesn't force it. Second, don't assume this overrides team-level restrictions that already existed. Nothing in GR 2026-002 stops an individual team from setting a stricter local policy. Third, don't treat this as Debian settling the AI copyright question. The resolution says explicitly that it isn't trying to. Fourth, don't read Haber's "let's revisit this in two years" comment as a sunset clause. There's no built-in expiration in the text, just an acknowledgment that the project might reopen the debate. And fifth, this only covers Debian's own packaging, documentation, and infrastructure. It says nothing about how upstream projects Debian merely repackages choose to use AI internally.
Tools that make this easier
Whichever way your own project lands on this, the habits Debian's policy assumes — reviewing AI output line by line, checking licensing before you submit, keeping sensitive data out of third-party models — are worth building regardless of the vote. My best AI tool for code roundup compares assistants on exactly the accountability and output-quality axes this resolution is arguing about, and my AI coding assistant guide walks through what to check before you trust one with a real contribution. If GitHub Copilot or ChatGPT is your daily driver for patches, my GitHub Copilot vs. ChatGPT comparison covers how each handles suggestions you still have to own. Before adopting any tool project-wide, my AI tool ratings hub is a reasonable gut check against vendor marketing. And if you want the contrast case, Oracle went the opposite direction earlier this month — see Oracle bans AI-generated code from OpenJDK for how a differently-governed open-source project answered the same question.
My take
Debian picked the option that changes the least in practice, and I think that's the right read of a community this split — 64/36 isn't a mandate for anything more aggressive in either direction. What I'd watch for is exactly the gap between what the policy assumes and what actually happens: optional disclosure means a maintainer reviewing a patch has no reliable signal that it was AI-assisted at all, which puts the entire "understand it before you submit it" requirement on the honor system. Haber himself seems to expect this isn't final — his own comment about revisiting it in two years reads less like confidence and more like a truce everyone could live with today. I'd bet the next real fight in Debian is about disclosure specifically, once someone ships an AI-generated patch nobody can explain.
Frequently Asked Questions
Did Debian ban AI-generated code?
No. The winning proposal, "Responsible Use of Generative AI," explicitly neither endorses nor prohibits generative AI tools. Two stricter ban-leaning proposals were on the ballot and both finished behind "None of the above."
Is AI disclosure mandatory under Debian's new policy?
No. Disclosure is encouraged but not required. A separate proposal from Lucas Nussbaum would have mandated a "Generated-By:" commit trailer, and it lost to Haber's option 148–203.
Who sponsored the winning proposal?
Debian developer Marc Haber sponsored Option E, "Responsible Use of Generative AI," which won General Resolution 2026-002.
What was the final vote margin?
Option E beat the second-place option 203 votes to 148, and beat the remaining options by wider margins, including 232–115 and 251–139 against the two most permissive and most restrictive alternatives still in contention.
Does this policy affect projects that Debian only packages?
No. GR 2026-002 is scoped to Debian's own packaging, documentation, translations, and infrastructure. It says nothing about how upstream projects Debian repackages choose to use AI internally.