Codeberg members ban hosting ‘mostly’ AI-generated code

Codeberg, the nonprofit alternative to GitHub used by number of popular FLOSS projects, will no longer host software that’s “mostly” AI-generated.

A vote closed yesterday (22 July, 2026), passing 358 to 144 with 14 abstentions on roughly 50% turnout to stop hosting software that’s largely vibe-coded. A second motion, also passed, commits Codeberg to never training AI models on hosted code or user data.

Not that it was ever likely to.

A Terms of Use amendment now prohibits projects that “mostly consist of code written by” generative AI tools.

In a blog post, Codeberg explains the economic need for a ban, explaining that SSDs used to cost the platform ~€700 a few years ago, but now run it to €3,700 each. The groundswell in vibe-coded projects is consuming its CI/CD and storage budge.

That’s unsustainable day-to-day for a service funded by donations. Codeberg was, like so many code sharing repos, built to enable humans to collaborate on free software together. Hosting single-use code that nobody wrote to be maintained runs counter to that.

Or, to quote the company, sharing the output of a prompt simply “does not make the world a better place”.

This is not a blanket ban – exceptions exist

Before anyone hosting code on Codeberg that has some traces of AI panics, there’s a get-out in that “mostly” AI generated code is now prohibited.

That ‘mostly’ is an important qualifier as it means this isn’t an outright blanket ban on all uses of AI. A human-driven project isn’t going to fall foul of the new rules by accepting the odd LLM-generated contribution (knowingly or otherwise), for instance.

There also will no be a mass purge, so existing projects with an active community or history predating LLMs won’t find their repos deleted overnight.

What aren’t safe are existing autonomous agent-built repos, projects whose resource burden isn’t justified by its contributor count and rojects whose code was never written by a person at all.

Codeberg may be named after something cold, but it isn’t in the business of being a deep freeze for code nobody wrote, nobody cares about and, chances are, nobody will ever look at again.

The move has echoes of Flathub’s recent policy change on AI-assisted code, which focused on the human cost of its review volunteers drowning in automated submissions of throwaway projects. Codeberg’s case is blunter: the literal cost of having to host it all.

And the “externalised cost” of LLMs and the push for AI-in-all-things is something we all feel when we look at hardware prices, energy costs and the environmental damage – irrespective of whether we use an LLM to one-shot another note-taking app the world doesn’t need or not.