Codeberg just voted to ban projects “mostly” written by generative AI tools. The result: 358 to 144, 71% in favor, roughly 50% member turnout. The Codeberg AI-generated code ban is already live in the Terms of Use. If you use Claude, Copilot, or Codex as the primary author on your repos, Codeberg is no longer your platform.
What the Policy Actually Says About AI Code
The updated Terms of Use are direct: “You must not share projects that mostly consist of code written by ‘generative AI’-tools (including services such as Claude, OpenAI Codex).”
Two resolutions passed at Codeberg’s annual assembly on July 22. The first bans AI-heavy repos. The second is a binding commitment that Codeberg will never use project or user data to train LLMs. Both passed decisively.
What’s allowed: ordinary projects where AI contributes occasionally, small experiments, established communities that accept some LLM-assisted contributions. What’s not: repos where AI wrote the majority of the code, regardless of how much human review happened afterward. Reviewing AI output does not create an exemption.
The “Mostly” Problem
Here’s where this gets complicated. Codeberg has not defined a percentage threshold for “mostly.” There is no detection methodology. Enforcement will be complaint-driven, handled by volunteers who are already stretched. Codeberg acknowledged enforcement details are still to be worked out separately.
Critics — including Armin Ronacher, creator of Flask and co-founder of AI firm Earendil — called it “a very bad move” according to The Register. Others are blunter: the policy is unenforceable as written. Nobody, including Codeberg itself, has defined what “mostly” means on a given project. That ambiguity will create inconsistent moderation. A developer who ships clean, AI-assisted code might get flagged based on project patterns — frequent releases, heavy CI/CD — that Codeberg associates with pure AI generation.
Why Codeberg Did This
The reasoning is economic, legal, and cultural — and all three are real.
Cost: SSDs that cost Codeberg €700 a few years ago now run €3,700. AI company crawlers scrape the platform aggressively, loading every issue filter variant and full commit history, hammering the database. The surge in AI-generated repos consumes CI/CD and storage budgets that donations barely cover.
Copyright: The legal status of AI-generated code remains unresolved. Codeberg also cited “license laundering” — copyleft requirements stripped when copyrighted training data gets regenerated through AI tools. Hosting such projects creates liability the non-profit cannot absorb.
Community trust: Codeberg put it plainly: “The widespread use of LLMs in FLOSS is becoming a multidimensional attack on the trust between contributors.” Ghost projects — heavy CI/CD pipelines, multi-platform releases, zero community — consume resources proportionate to a well-maintained team project but have no actual maintainers. Furthermore, they burden reviewers and create an uneven playing field against human-maintained projects.
The Broader Open Source AI Divide
Codeberg is not alone in wrestling with this. Debian’s ongoing General Resolution is still debating three competing proposals on LLM usage — but unlike Codeberg, Debian hasn’t resolved anything yet. Meanwhile, the open-weight AI coalition letter signed by NVIDIA, Microsoft, and Meta pushed in the opposite direction, arguing for open model distribution.
There’s an irony worth noting. In 2025, the Zig project left GitHub for Codeberg specifically because of concerns about GitHub’s AI integrations. Codeberg attracted developers who were skeptical of AI in open source. Now Codeberg is the platform enforcing that skepticism as policy — which may be exactly the point, or may alienate developers who moved there for entirely different reasons. See also Hackaday’s breakdown of both the crypto and AI bans for the full governance picture.
What Developers Should Do Now
If you build open-source projects with AI as a significant tool, your options have clarified:
- Self-host Forgejo — Codeberg runs on Forgejo. The engine itself carries no such restriction. You control your own instance and your own terms.
- Gitea — Similar situation, no AI policy.
- Stay on GitHub or GitLab — Both are actively integrating AI features and neither is restricting AI-generated code.
Codeberg’s ban is effectively a market segmentation event. The platform has chosen a lane: human-authored free software, community-governed, donation-sustainable. That is a coherent product decision. However, “mostly AI” is not a technically enforceable standard — it is a value judgment. Moreover, value judgments applied inconsistently across a volunteer moderation team will create friction for legitimate developers who use AI as a tool, not a replacement for judgment.
The practical move for serious open-source developers: self-host on Forgejo, keep full control. Codeberg’s governance will keep evolving, and today’s “mostly AI” ban sets the precedent for tomorrow’s policy expansions. The question is not whether the ban was the right call — reasonable people disagree. The question is whether an undefined “mostly” standard can be enforced fairly. Based on the evidence so far, the answer is not yet clear.
