Codeberg, the nonprofit code forge a lot of developers have been quietly moving to as a GitHub alternative, just voted to ban vibe-coded projects, meaning repositories that mostly consist of code written by AI rather than a human. The vote passed by a wide margin. The reaction to it hasn’t been nearly that clean.
The Quick Version #
What happened: Codeberg e.V. members voted to ban projects that “mostly” consist of code written by generative AI tools.The vote: 358 in favor, 144 against, 14 abstentions, with roughly half of all active members turning out.Who gets named: The new rule specifically calls out tools like Claude and OpenAI Codex.The catch: Nobody, including Codeberg itself, has defined what “mostly” actually means.What’s not happening: No mass deletions. Moderators say they’ll handle it case by case.
What Actually Got Voted On #
Two AI-related motions went up during the annual assembly of Codeberg e.V., the German nonprofit association that runs the platform, and both passed.
The first is simple. Codeberg promises it will never use anyone’s code or user data to train an LLM. That matches what the org already says in its privacy policy, so almost nobody pushed back on it.
The second motion is the one blowing up right now. It adds a new line to Codeberg’s Terms of Use banning projects that mostly consist of code written by generative AI. After a two-week vote among dues-paying members, it passed 358 to 144, with 14 abstentions. Roughly half of all active members turned out, and about 71 percent of everyone who took a side voted yes.
Codeberg has grown enough over the past couple of years to host recognizable projects like KaOS and Readeck alongside thousands of smaller ones, so a rule change like this affects real people, not just a hypothetical.
The New Rule, Word for Word #
Here’s what actually got added to the Terms of Use. Members “must not share projects that mostly consist of code written by” generative AI tools, and the rule names Claude and OpenAI Codex specifically as examples.
The justification isn’t really about code quality. It’s about copyright and safety. AI-generated code sits in a legally unclear spot on ownership, and heavily AI-written contributions tend to get less real scrutiny before merging, which leaves more room for bugs or malicious code to slip through untested.
Why Codeberg Says It Had This Coming #
The rule change shipped alongside a much longer blog post called “Protecting our FLOSS commons from LLMs,” and the reasoning goes well past copyright.
Hardware got brutally expensive
Storage prices are a big part of it. Codeberg says the type of drive it used to buy for around 700 euros a few years ago now costs roughly 3,700 euros, when it’s even available. The org owns its hardware outright, so it isn’t exposed to cloud pricing hikes directly, but replacing or expanding that hardware costs dramatically more than it used to.
The crawlers won’t stop
Then there’s the crawler problem. Codeberg says AI companies scraping the site for training data try to load every page combination that exists, including every issue filter variant and every historical version of every file, whether or not anything in it ever changed. That traffic burns through database resources and forces the team to build rate limits and blocks that end up punishing real users right along with the bots.
Ghost projects are eating the budget
There’s also what Codeberg calls ghost projects. Single-developer repositories with almost no users are burning through as much CI/CD and storage budget as some of the platform’s biggest actual community projects, complete with heavy test suites and frequent, oversized release builds that almost nobody downloads.
Trust is the real casualty
Maintainers are stuck reviewing a growing pile of low-effort, AI-generated pull requests, and it’s getting harder to tell which projects are run by experienced developers and which ones are mostly AI output with no real oversight behind them.
Codeberg raises a copyleft-specific worry too. Code released under a reciprocal license can effectively get laundered clean once a model regenerates something close to it, stripping out the obligations the original license carried.
Some contributors are also getting wrongly accused of using AI when they didn’t, while others coach their tools to hide the tells, which just pushes everyone toward reviewing each other more suspiciously. Codeberg argues this drags the whole ecosystem in the wrong direction. Instead of contributing to and improving tools other people already maintain, more developers just prompt a one-off version of whatever they need from scratch, so total code output goes up while the number of things anyone is actually maintaining goes down.
The environmental and civic angle
There’s a civic cost too. Communities near AI-focused data centers are already dealing with higher electricity and water prices, along with more air and noise pollution, and the companies running those centers are lobbying to loosen environmental rules so they can keep expanding, in some cases turning to fossil fuels to meet demand. Codeberg points to Frankfurt as an example, saying data centers there already account for 40 percent of the city’s electricity use, with demand still climbing.
What’s Fine, and What’s Probably Not #
Codeberg was clear that this isn’t a mass purge. There’s no plan to scan every repository and build a takedown list. Moderators say they’ll handle reports case by case as they come up, going by a few early, informal guidelines.
Likely to be fine
- Projects with an active community actually maintaining the code
- Projects with real history that predates heavy AI use
- Maintainers who unknowingly accepted an AI-written contribution from someone else, as long as the project doesn’t otherwise lean on AI heavily
Discouraged, but tolerated in practice
- Small side projects and experiments that use very little in the way of resources
- Personal tools and scripts that were never going to attract a community anyway
Likely not welcome
- Projects generated autonomously by AI agents
- Projects heavily written and maintained using AI
- Projects whose resource use is way bigger than what the people actually involved could plausibly have built by hand
- Tooling built specifically to make using other AI tools easier
- Repeat contributors who keep ignoring a project’s own stated anti-AI policy
Nobody Can Agree on What “Mostly” Means #
This is where most of the actual argument is happening, and Codeberg’s own contributors raised it before the vote even closed.
The Terms of Use never define “mostly.” Not as a percentage, not as a method, nothing. Is a project mostly AI-written at 51 percent? 80 percent? Does it get measured by line count, by commit count, or by how much a human actually understood before hitting merge?
When contributors pushed on this during the review period, the response was fairly blunt. The wording matches how Codeberg already handles other banned content on the platform, and there wasn’t time left to rewrite it before the vote closed. That answer satisfied almost nobody in the comments.
Enforcement is the other open wound. Since Codeberg says it won’t be actively scanning projects, most of this will likely come down to reports and moderator judgment calls rather than anything measurable. Several critics have already pointed out the obvious irony: catching AI-generated code with any real consistency would probably require another AI tool to do the detecting.
The Community Reaction Is Split Right Down the Middle #
Reaction split fast. It started right in the pull request where Codeberg’s own contributors debated the wording, then spilled over onto Reddit and Hacker News once the vote results went public.
Supporters are mostly just tired. Years of reviewing low-effort, AI-assisted pull requests has worn a lot of maintainers down, and they see this as a filter, not a ban on the technology itself. The general argument goes that using a tool responsibly and actually understanding your own code is fine, and dumping a project you don’t understand onto other people to review is not.
Critics are worried about something different. A vague rule enforced case by case can be applied inconsistently, and “mostly AI” will always be a judgment call rather than something you can check objectively. Several developers said they’re already planning to self-host on Forgejo or Gitea instead, specifically to sidestep a policy that might land differently depending on who happens to review their project.
The reaction emojis on the merge announcement itself skewed far more positive than the formal vote did, with more than a hundred thumbs up against fewer than a dozen negative reactions, hinting that the loudest voices online lean more supportive than the actual paying membership.
The Copyright Fight Underneath All of This #
Part of the disagreement traces back to a genuinely unresolved legal question. Who actually owns AI-generated code in the first place?
Some commenters argued that if AI output isn’t copyrightable at all, it effectively lands in the public domain, meaning anyone can already do whatever they want with it. Others pushed back that a total lack of ownership creates its own mess, since code with no license attached can’t be safely combined with copyleft projects or protected if someone changes the terms on it later.
Nobody landed on a clean answer, and the underlying law genuinely differs by country. One commenter pointed to the ongoing New York Times lawsuit against Microsoft and OpenAI over training data as proof the bigger question is far from settled anywhere.
What Happens Next #
For now, Codeberg says the immediate impact is small. No mass deletions, no automated scans, no retroactive digging through anyone’s commit history. The org says it will keep using moderation calls, and its own follow-up posts, to clarify the gray areas as they actually come up. Codeberg isn’t the first FOSS project to draw a line here either. Commenters pointed to Godot’s team reportedly limiting LLM use internally, including for contributor communication, as a sign more projects are heading the same direction.
The bigger story isn’t really about Codeberg specifically. Every code forge is going to have to answer some version of this question eventually. What do you do when a growing share of the code showing up at your door was written by something other than the person submitting it? Codeberg just happened to be the platform whose members voted on an answer first.