Debian opened voting on August 15 on a General Resolution covering the use of large language models within the project. Voting closes at 23:59 UTC on August 28, and only Debian Developers can cast a ballot.

The ballot lists eight proposals plus “None of the above.” It started smaller and grew during the discussion period, which began on July 24, as developers proposed amendments rather than settling behind existing options.

The range of positions

The eight choices span most of the reasonable positions on the question:

• Choice 1 amends the Social Contract with explicit language: “We will not allow direct contributions to Debian written with the use or assistance of large language models (LLMs) or other generative AI tools.”
• Choice 3 rejects LLMs “as far as practical” and puts the rule in the Code of Conduct rather than the Social Contract.
• Choice 4 accepts AI contributions for Debian-specific work.
• Choices 2, 5, and 6 are responsible-use and cautious-approach variants.
• Choice 7, titled “Debian is created by humans,” requires a human author without dictating which tools that human uses.
• Choice 8 skips the code-quality argument entirely and objects on environmental grounds.

The ballot math matters more than the rhetoric

Here is the detail that decides this: Choice 1 amends a Foundation Document, which requires a 3:1 supermajority. Every other option on the ballot needs only a simple majority.

Debian uses a Condorcet method with ranked ballots. In that system, an asymmetric supermajority requirement is close to decisive. An outright ban has to clear a bar three times higher than any of its competitors, and several of those competitors are close enough to it in spirit that developers uneasy about AI can rank a milder option second without feeling they have abandoned the position.

An outright ban passing is not impossible, but it requires overwhelming rather than majority support.

What is actually being argued about

The tension is real and it is not primarily about whether AI writes good code.

On one side: reviewer burnout. Debian runs on volunteer review capacity, and a flood of machine-generated patches consumes that capacity regardless of patch quality. The Linux kernel’s networking maintainers have described themselves as “completely overwhelmed” by exactly this during the 7.3 merge window. There is also a genuine copyright question about the provenance of model-generated code entering a distribution that takes licensing seriously enough to have a Foundation Document about it.

On the other side: many maintainers already use these tools daily, and a ban that cannot be enforced mostly produces a rule people quietly ignore, which is worse for a project’s culture than having no rule.

Why this is worth watching outside Debian

Debian is not the first project to argue about this, but it is one of the first large ones to put it to a formal, binding vote with a documented process and a public ballot.

Whatever wins will become the reference point other projects cite, in the same way Debian’s non-free firmware vote became the reference for that question. The OpenSSH project’s decision to ship security releases more frequently because of AI-assisted vulnerability reports is the same underlying pressure showing up in a different form.

The result should be published shortly after voting closes on August 28.

Update, August 29: The vote has resolved. Option 5, “Responsible Use of Generative AI,” won, and the Social Contract amendment banning LLM-assisted contributions failed to reach its supermajority. See Debian Votes for Responsible Use of Generative AI for the margins and the policy text.