Codeberg members ban 'vibe-coded' projects and pledge no LLM training on user data

Codeberg, a nonprofit and member-run host for free and open source software, asked its roughly 700 active members a yes-or-no question, and got an answer that breaks the usual pattern of platform AI policy. The members voted, 358 to 144 with 14 abstentions on a turnout near fifty percent, to amend the platform's terms of service so that "vibe-coded" projects, meaning code where a large language model does most of the writing and a human reviews rather than authors, can no longer be hosted there. They passed a second motion at a similar margin pledging that Codeberg itself won't use hosted repositories or user data to train generative AI.

The pair of motions, ratified in a 14-day asynchronous vote that closed the day before the Codeberg announcement, is the first time a registered free-software association has put both a no-train pledge and a project-acceptance bar against AI-dominant code to an actual membership ballot. About one in three active members voted against the more restrictive motion, and roughly half the eligible members didn't vote at all. The community itself knows the work isn't finished.

The two motions don't change as much as the headlines will suggest. The no-train pledge is a privacy and data-use commitment. The operator, the registered association Codeberg e.V., won't pipe hosted repositories into model training. It isn't a technical scraping block, and the post doesn't claim one. AI companies that want to crawl Codeberg for training data still have to be deterred at the network layer or by robots.txt, and the post announces no infrastructure changes. The vibe-coding motion is a ToS line: a project-acceptance rule that maintainers will have to enforce by review, with the hardest moderation questions left open.

SourceHut founder Drew DeVault has been making the cost-externalization case against LLM-driven crawling for years, and Codeberg's framing borrows that vocabulary directly. The argument is that GPU bills, energy use, and environmental footprint of model training land on people and infrastructure that didn't ask for them, including nonprofit forges whose bandwidth and storage are donated and capped. It's a policy claim with directional force, not a measured figure. The post names the cost categories, not a number, and a reader who wants an emissions accounting won't find one.

The moderation problem sits inside the word "vibe-coded." A project whose commits are mostly AI-generated with light human review is the easy case. The hard cases arrive immediately: a documentation translation drafted by an LLM and checked by a human; a dependency upgrade produced by a coding agent and merged after a one-line review; a research-paper-to-code conversion where the human is the experimenter and the model is the typist. Someone on Codeberg's project-review side has to draw the line. The post doesn't publish the line.

The Zig programming language, a major Codeberg-hosted project, has a community-attributed hostility to LLM agents in its contribution policy, though the supplied receipts don't include a primary Zig-side statement. That's the kind of signal the new ToS will reward. Projects whose maintainers were already skeptical of AI-author contributions now have a platform whose rules agree with them. Projects whose maintainers rely on AI tooling for routine work will read the same ToS as a narrowing of where they can host.

Reaction in a Hacker News thread splits along a familiar fault line. Some commenters read the vote as principled governance, a small community deciding what its shared infrastructure is for. Others read it as a brand-identity move, a nonprofit staking out a position in the broader open-source-AI debate to differentiate from GitHub. The two readings don't exclude each other, and XDA and explainx.ai cover it the same way. The vote margin is the evidence that the membership itself wasn't unified.

Codeberg's maintainer corps now has to publish a review rubric, and someone has to staff it.