Skip to content

Policy & regulation

Ban on LLM-assisted code goes to a Debian vote

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: thenewstack.io

Debian Developers have until 23:59:59 UTC on August 28 to rank eight proposals on whether large language models belong in the distribution’s own work. The strictest, Proposal A from Matthias Geiger, would ban direct contributions written “with the use or assistance of large language models (LLMs) or other generative AI tools,” The New Stack reported.

Geiger argues from Debian’s “reputation for stability” and ties widespread LLM use to the “move fast and break things” attitude: “While [that approach] is common in many parts of this industry, it is contrary to what makes Debian Debian, and is inappropriate for Debian contributors.”

The ban would cover Debian source packages and other software developed by Debian, official Debian web resources, packaging work, native tools such as the lintian package checker, and documentation and translations written by Debian contributors. Upstream projects that use generative AI stay outside it, and Debian could still take patches and security fixes from upstream whether or not AI was involved in writing them. A maintainer could package a third-party AI coding tool but could not use it for new work on Debian itself.

The proposal text on Debian’s vote page gives two main arguments. The first is copyright. LLM output has “very unclear legal status”, while Debian Policy and the DFSG, the Debian Free Software Guidelines, demand complete clarity on licensing. Human-written contributions with unclear status are already refused, so the proposal says LLM output should not get an exception. The second is quality. An LLM “merely produces syntactically likely combinations of the training data”, the proposal says, so a generated package would mix material from across the archive’s history: watch files that don’t work, overrides out of context, imaginary copyright. And a new contributor would not know how to fix it. The text also notes that it was “written organically, without LLM assistance.”

Enforcement would rest only on “good faith compliance”, and the proposal admits there is no detection mechanism. Matt Brown, principal AI architect at Endor Labs, makes that gap his main objection. “Banning AI-assisted code doesn’t eliminate AI-assisted code – it’ll just make its use harder to disclose – what’s stopping someone from copy/pasting everything, or even manually typing out what it generated?” he told The New Stack. Brown says he understands the worry about unreviewed, low-quality contributions in a project built on stability, but a blanket ban “risks confusing the tool with the quality of the work.” His alternative is transparency, human accountability, rigorous testing and the same review bar for every contribution.

Lucas Nussbaum’s Choice 2 allows LLM contributions under six conditions. Ian Jackson’s Choice 3 moves the rules into the Code of Conduct, keeps person-to-person messages human-only and makes disclosure mandatory. Pierre-Elliott Bécue’s Choice 4 sets out a responsibility model, and Marc Haber’s Choice 5 says Debian “neither endorses nor prohibits.” Gard Spreemann’s Choice 7 bans AI output but not AI tools. Two more options are a statement under the pledge “Debian is created by humans” and one titled “avoid the use of LLM: climate destruction is a deal breaker.” Jackson also appears among Proposal A’s seconders, along with Bécue, David Bremner, Simon Richter and others.

Voting opened on August 15, after a discussion period that began on July 23. Developers vote by GPG-signed email in a secret ballot. They rank nine options, the eight proposals plus “None of the Above”, and the ranking is counted by the Condorcet method, which compares every option against every other one instead of counting first preferences. Proposal A, framed through Debian’s social contract, needs a 3:1 majority; the other seven need a simple majority.

Joe Phillips, author of the Hybrid Workforce Standard, sees less disagreement than the ballot suggests. In his reading, all sides agree that code provenance must be declared and are arguing about “how big the sticker should be.” “That is not a war over principle – that is a formatting dispute with a manifesto attached,” he told The New Stack. Phillips says the strongest argument sits in the community section: “contributors who lean on a model never learn packaging, so they can never replace the maintainer who burns out.” He calls that deskilling the one harm that compounds, but says a ban teaches nobody anything. “If Debian cares about its pipeline, the answer is apprenticeship, not prohibition: make new contributors work in shadow mode, review them like juniors, and let them earn authority.”

Justin Beals, CEO and founder of the compliance platform Strike Graph, calls the ban “the wrong fix for the actual problem.” “Debian’s ban isn’t really about banning AI,” he said. “It’s an admission that nobody has built a reliable way to verify what an AI agent actually produced before it lands in a codebase this many systems depend on, including infrastructure running in orbit.” He argues that any project running on trust and reputation is now exposed to contributors who can generate endless plausible-looking code, and that a ban removes one way of triggering that problem while leaving the trust model untouched. “Banning LLM contributions treats the symptom, not the disease,” Beals said. The projects he expects to do well will build provenance tracking, review depth tied to actual risk, and evidence of meaningful human validation into the contribution pipeline.

Later, FOSS Linux reported that when the vote closed on August 28, Haber’s Choice 5, “Responsible Use of Generative AI”, won. It said 1,039 active developers had been eligible to vote and that Proposal A “didn’t even come close” to its 3:1 supermajority.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.