Codeberg members just voted to ban AI trained on their code and restrict 'vibe coded' projects on the platform

[Pale Connor]

Members of Codeberg e.V., the German non-profit behind the Codeberg code hosting platform, an open source alternative to GitHub, voted at their annual assembly to formally commit the organization to never using code or user data hosted on the platform to train generative AI tools, and separately to restrict what the organization calls vibe coded projects, meaning software built largely or entirely through LLM generated code with minimal human authorship or oversight. Both motions passed after a 14 day asynchronous voting window following live discussion at the assembly

The AI training rejection formalizes a position Codeberg's privacy policy already implied, turning what was previously an informal stance into an official association commitment. The vibe coding restriction proved far more contentious, passing 358 to 144 with 14 abstentions, meaning nearly three quarters of voters backed it but a genuinely sizable minority did not, and roughly half of Codeberg's active membership turned out to vote, a rate the organization itself flags as unusually high for this kind of decision

Codeberg ties both votes to a concrete infrastructure cost problem it says has been building for a while, AI companies operating web crawlers that hammer the platform attempting to load every single page individually, including every issue filter variant and every historical file state at every past commit, generating far more database load than simply cloning a repository through normal Git operations ever would. The organization says defending against this traffic has consumed system administrator time that could otherwise go toward other infrastructure work, and that the same rate limits and access blocks built to slow down aggressive crawlers end up degrading service for legitimate users too

The reasoning behind the vibe coding restriction specifically is more speculative and Codeberg is fairly upfront about that, maintainers say they've noticed a pattern where a single developer's LLM assisted project runs CI/CD pipelines sized for a full team, ships frequent heavy release binaries, and claims platform support far beyond what its actual user base would seem to warrant, consuming server resources as though a real community existed behind it when in practice one person is prompting a model that converts compute into commits. Codeberg has not published data substantiating how widespread this pattern actually is across its hosted repositories, and has promised a separate follow up post detailing exactly how the new restriction will actually be enforced, a genuinely open question for any maintainer currently hosting an LLM assisted project there

Codeberg members vote to reject LLM training and vibe codingwww.developer-tech.com