Codeberg Bans Projects Mostly Written by Generative AI

Codeberg, the German nonprofit code-hosting platform, now forbids hosting projects that “mostly consist of” code written by generative AI tools. The clause was merged into the platform's Terms of Use in the early hours of today after a membership vote, and it names Claude and OpenAI Codex directly as examples. This is a different kind of rule from the AI bans that open-source projects have been announcing all year: those block your pull request, this one blocks your repository.

• The rule is live, not proposed. It sits in Codeberg's Terms of Use as § 2 (1) 7, effective now, after a proposal opened on 29 June 2026 passed a vote of the association's membership.
• The stated reason is copyright, not code quality. Codeberg argues such projects have “unclear copyright status” and therefore cannot satisfy the platform's requirement that everything hosted be validly licensed as free and open source.
• The threshold is undefined. “Mostly” is not quantified anywhere in the text, and contributors in the discussion pushed unsuccessfully for stricter wording.
• A second ban passed in the same cycle. The same assembly amended the Terms of Use to name cryptocurrency projects as content that harms Codeberg's reputation.

What exactly does the new rule say?

The added clause is short. Under § 2 (1), the section listing what you agree to when posting content, Codeberg inserted a seventh item: you must not share projects that mostly consist of code written by “generative AI” tools, including services such as Claude and OpenAI Codex. The justification given in the clause itself is that such projects have an unclear copyright status, and that they carry little safeguard against including harmful code.

RelatedMicrosoft Open-Sources Comic Chat After 30 Years

Two things are worth noticing about the wording. First, it targets projects, not contributions, which is what makes it a hosting rule rather than a contribution rule. Second, it explicitly names two commercial products. Terms of Use documents usually stay vendor-neutral, and naming Claude and Codex reads as a deliberate signal about which category of tool the association has in mind: autonomous code generation, not autocomplete.

Why did Codeberg tie this to copyright instead of quality?

This is the part most of the reaction has skipped, and it is the load-bearing argument. Codeberg's Terms of Use already require, in § 2 (1) 1, that public repository content carry a licence granting everyone the rights to use, copy, study, modify and redistribute it. § 2 (1) 3 goes further: you may only share content you actually have the legal right to share under those terms.

Those two clauses predate the AI debate, and together they create the trap. To license code under the GPL or MIT, you need to hold copyright in it. If the copyright status of largely machine-generated output is unsettled, then the person uploading it cannot cleanly grant the rights that Codeberg's licensing rule demands. The new clause does not invent a new principle so much as apply an existing one, which is legally tidier than a rule that says AI code is bad code.

The second justification, the harmful-code point, hooks into § 2 (1) 5, which bars repositories containing code that can run accidentally during a normal development workflow. The argument is that a project nobody meaningfully reviewed has weaker safeguards against that than one a human wrote line by line.

How is this different from Godot or GNOME banning AI code?

Several projects moved first this year. Godot rewrote its contribution policy on 1 July to prohibit contributions made entirely by AI and to require disclosure when AI assisted. GNOME barred AI-generated submissions from its Extensions repository. QEMU declines contributions believed to derive from AI-generated content, and NetBSD treats LLM output as presumed tainted code requiring written approval before it can be committed.

Every one of those is a project deciding what it will accept into its own tree. Codeberg is a host deciding what it will store. If Godot rejects your patch, your fork still lives wherever you put it. If Codeberg applies § 2 (1) 7, the repository comes down.

What happens if a project gets reported?

The enforcement path was already written into § 2 (2) and now applies to this clause too. A first failure to comply leads to immediate removal of the content along with a warning. Further violations may result in immediate account suspension. In cases that are not obvious, and an accusation that a repository is mostly machine-written will rarely be obvious, the decision falls to the presidium and requires a simple majority.

That is the honest answer to the detection problem, and it is worth being clear that it is a governance answer rather than a technical one. There is no reliable detector for machine-written code, and Codeberg has not claimed to have one. What it has is a committee vote. Whether that scales depends entirely on how many disputes arrive.

Contributors raised exactly this during the discussion. One argued that “mostly” invites a lenient reading, asking whether a project that is 49 percent generated would be fine, and suggesting “in significant part” instead. Questions about whether the rule reaches private repositories, and what “sharing” covers, were acknowledged and deferred as out of scope. The clause shipped with the loose wording intact.

RelatedPython 3.14 makes free-threading real: the GIL era is ending

How did this land, and what else passed?

The proposal was opened on 29 June 2026 by Bastian Greshake Tzovaras under the title “prohibit LLM-extrusions,” went to the association's 2026 assembly, passed a membership vote, and was merged today by the maintainer known as Gusted. It did not travel alone. A second proposal from the same cycle, opened on 2 July, amended the list of content that harms Codeberg's reputation to name cryptocurrency-related projects explicitly. Both edits reached the live Terms of Use within hours of each other.

1. 29 Jun 2026Proposal opened to prohibit projects mostly written by generative AI Filed for the 2026 assembly
2. 1 Jul 2026Godot bans contributions made entirely by AI Project-level, not platform-level
3. 2 Jul 2026Second proposal filed to disallow cryptocurrency projects Same assembly cycle
4. 22 Jul 2026Both clauses merged into the Terms of Use and live Membership vote passed
5. NextThe presidium interprets “mostly” in the first contested case Sets the real threshold

Who does this actually affect?

Codeberg is small next to GitHub, but it has spent the past year absorbing projects that left precisely over AI. Gentoo announced its presence there in February 2026 citing pressure to adopt Copilot. Dillo's founder moved for similar reasons. Zig, whose 1.0 work we covered on 12 July, develops there too. For that population, this rule is the platform confirming the values they moved for, which is why the loudest objections came from inside rather than outside.

The practical exposure is narrower than the headline suggests. If you use a model for completion, refactoring or tests inside a codebase you designed and understand, nothing here touches you. The clause is aimed at repositories that are essentially model output with a README, the category maintainers have started calling slop. The uncomfortable middle is the developer who prompted an agent through an entire weekend project and genuinely cannot say whether the result is 40 or 80 percent generated.

Our take

The copyright framing is the smart move and the vague threshold is the weak point, and both matter more than the culture-war reading this story will get. By anchoring the ban to an existing licensing requirement rather than to a judgment about quality, Codeberg built a rule it can defend on legal grounds without having to prove anything about how good the code is. That is a much stronger footing than “we do not like AI.”

But a rule that cannot be measured is a rule enforced by complaint. With no definition of “mostly” and no detection method, what determines outcomes is who gets reported and how the presidium feels about that specific case. Expect the first contested removal to matter far more than the text itself, because it will establish the working threshold that the clause declined to write down. The genuinely new thing here is not that a community dislikes generated code. It is that a host has made the origin of your code a condition of storage, and other platforms now have a template to copy.

Original analysis by GenZTech, based on Codeberg's published Terms of Use and the governance thread.