Codeberg's members to vote on banning repos 'mostly' made of AI-written code

The non-profit, member-governed code host Codeberg is asking its members to vote on a Terms of Use extension that would add a new forbidden category: "LLM-extrusions," or projects that mostly consist of code produced by an AI model (PR #1253).

The proposed text borrows the same "must not share" line that already polices hate and violence content in the Codeberg Terms of Use, so the legal mechanic is an acceptable-use extension rather than a new code-of-conduct track. Anyone who has read the ToU already knows the shape of the rule; the only new word is "mostly." The proposal was agreed between gedankenstuecke and Gusted after the previous Assembly and is being sent to the membership ballot, with minor board edits possible if it generally passes (commit 96fac426).

The live disagreement on the PR is about scope. Reviewers are asking whether "sharing" covers pull requests into other Codeberg repositories, whether private repositories are in scope, what "mostly" means in practice (more than half the files? most of the lines?), and whether a small utility that someone reads by hand and then re-shares would still count as an LLM-extrusion. The thread is open on the PR page (PR #1253). At least one commenter, infosec812, has argued the copyright dimension is sufficient on its own, since LICENSE compliance is often broken regardless of how the code was produced. That pushback has not been dismissed; the proposal is still in review.

Codeberg is small, member-governed, and runs on community infrastructure. Codeberg e.V. is a registered German non-profit; the platform is a Forgejo-based alternative to GitHub, which gives the ToU action a governance and jurisdiction weight a commercial host would not carry. A member-run non-profit can move a rule like this faster than a corporate platform because there is no separate product, advertising, or developer-relations lane to coordinate with. The Assembly vote is the rule. Until then, the proposal is a workshopped text the board can still lightly edit if the membership generally approves.

Hacker News readers are already framing the move as a sibling to a separate resolution on "vibe coded" projects, though the two artifacts are distinct. Thread 49003386 headlines the vibe-coding ban, which is PR #1254, not the LLM-extrusion ToU now going to vote (HN 49003386). Thread 49003465 treats them together, with commenters weighing in on both the rule and its scope (HN 49003465). The two resolutions will likely be voted on in the same Assembly, but they are separate legal instruments with different rule shapes.

The non-profit's forum thread on the resolution is members-only at forum.codeberg.org/d/139, so the deeper "why" of the proposal, including any prior failed attempts at AI policy, is not on the public record. For an analysis, that gap is real: the organizational rationale behind the rule is in a venue the public cannot read. The 2026 Assembly date is the procedural clock, and the diff and the in-thread comment record are the rest of the reader's evidence base (full diff).

The rule is both too narrow to matter and too soft to enforce. Codeberg's user base is small, the text is still workshoppable, and the "mostly" threshold is undefined. None of that makes the rule meaningless; it makes the rule experimental, and it makes the Assembly vote the moment the workshop ends. A member-governed non-profit choosing to write AI-code policy by reusing its hate-speech clause is a named instance of how this corner of the open-source ecosystem is drafting its own answer, on its own timetable, with its own open questions still on the table.