Debian Developers Vote on LLM Use: Five Proposals Spanning Ban to Pragmatic Acceptance
The Debian project has opened a General Resolution vote on the use of Large Language Models by contributors. Discussion began on July 24, 2026, and the ballot will follow shortly after. Five competing proposals are on the table, ranging from an outright ban on AI-assisted contributions to a fully permissive framework with guardrails.
This is the latest installment in Debian's ongoing negotiation with AI. The project has been wrestling with this issue for years. A 2024 discussion kicked off the conversation, followed by a withdrawn resolution in 2025 that acknowledged the complexity of the topic. Early 2026 saw another draft that was set aside in favor of case-by-case handling. Now, the dust has settled into a formal vote.
Keep in mind that the scope is limited to Debian-specific work. Upstream projects, including those that might use AI internally, are off-limits for this resolution. Debian can only regulate its own packaging, documentation, and infrastructure.
The Five Paths Forward
Proposal A, led by Matthias Geiger, proposes a complete ban. It would forbid LLM-assisted source code, web resources, and official communications. The argument rests on copyright clarity, quality control, and ethical concerns about web scraping. It's the most restrictive option, though enforcement is acknowledged as difficult.
Proposal B, by Lucas Nussbaum, is the most permissive. It allows AI use under six conditions, including tooling terms compatibility, contributor accountability, and disclosure via Git trailers. Nussbaum's proposal has the strongest seconding with nine seconds, suggesting broad support for a pragmatic approach.
Proposal C, from Ian Jackson, takes a middle ground. It asks contributors to avoid LLMs "as far as practical" and makes AI use in messages to humans a Code of Conduct violation. Jackson highlights the inherent tension by noting that contributors with limited English proficiency may need AI for their native language.
Proposal D, by Pierre-Elliott Bécue, focuses on Debian-specific work and requires contributors to understand and defend AI-assisted code. It emphasizes personal responsibility and disclosure in commit messages, though light tools like tab-completion are exempt.
Proposal E, from Marc Haber, is the most measured. It neither endorses nor prohibits AI use, emphasizing that contributions must meet the same standards regardless of tooling. Disclosure is encouraged but not required, and confidential information must not be sent to third-party AI services.
The Enforcement Problem
All proposals grapple with enforcement. Proposal A admits it could be a challenge. Proposals D and E rely on good faith and self-assessment. Only Proposal C attempts to make AI rules binding through Code of Conduct violations. This enforcement paradox is a key theme in the debate.
The community is split between protecting volunteer reviewers from burnout and acknowledging the productivity gains of AI tools. Many maintainers already use AI and find it helpful. Banning it entirely could drive the practice underground or frustrate contributors who rely on these tools for accessibility.
Notably, Proposal A was written without language model assistance, a meta-commentary that has drawn attention. It's a statement of intent that underscores the ethical concerns about AI's impact on the ecosystem.
Voting and Next Steps
Debian uses preferential voting, allowing developers to rank all five proposals. This avoids spoiler effects and determines a Condorcet winner that best reflects community preferences. All active Debian Developers will receive a ballot via signed email.
Head here to the official vote page for full details on each proposal. The discussion period is still open, and developers are encouraged to engage in the debate before the voting phase begins.
This vote will shape Debian's approach to AI for years to come. The decision will reflect the project's values and priorities as it navigates an evolving technological landscape. Keep an eye on the results as they're announced.
