Codeberg's New AI Policy Sounds Simple. The Word 'Mostly' Is the Problem.

Codeberg, a non-profit open source code-hosting platform that bills itself as a community-governed alternative to GitHub, updated its terms on 22 July 2026 to bar projects whose code is "mostly" written by generative AI. Members of Codeberg e.V., the German association that runs the service, approved the change 358 to 144, with 14 abstentions, on roughly 50% turnout. The vote also passed a second motion: Codeberg will not train models on hosted code or user data, a commitment the platform says its privacy policy already covers and treats as largely symbolic.

Enforcement of the new rule now rests on whoever interprets the word "mostly" against a given project.

Codeberg's own case, laid out in its blog post, is economic rather than ideological. By Codeberg's own figures, the per-SSD cost on its storage fleet has climbed from roughly €700 a few years ago to about €3,700 now. At the roughly 1.10 USD-per-euro rate prevailing in July 2026, those figures translate to about $770 and $4,070 per SSD. The association blames the spike on "vibe-coded" CI/CD loops and repositories bloated by LLM-generated commits. The ToU amendment, in this reading, is a load-management rule dressed as a values statement. Codeberg acknowledges the "mostly" threshold is subjective, and the rule does not try to fix that: the policy is not a bright line. It is a permission slip for whoever applies it.

Armin Ronacher, the Python developer behind Flask and Jinja, read the rule the same way. He argues the word "mostly" cannot be enforced against a working codebase without splitting the diff line by line, and predicts the middle ground will be settled by community social pressure against borderline projects rather than by the terms themselves. Ronacher is not anti-LLM. He writes that LLMs done well could be tools that "reclaim power from large corporations and institutions," and he is explicit that Codeberg has every right to set its own rules. His critique is about clarity: a policy that hands discretion to moderators is not a ban and not an allowance. It is a delegation.

GitHub, the dominant code host, has so far chosen predictability over values signaling and has not changed its ToU to bar AI-generated code, a stance Ronacher's column credits for beating Codeberg's "democratic ambiguity" on infrastructure reliability. Codeberg is the first major host of any kind to put the question of "what is AI-written" into its terms, and the first to do it with a word that admits it cannot be enforced mechanically. Reporting from OMG Ubuntu and Explainx confirms the vote and the storage-cost figures but does not engage the enforceability question; that is left to Ronacher and to Codeberg's own admission in the post.

The community response, captured in the Hacker News thread on the news, is split along the lines the policy invites. Some commenters praise the stance; others argue that moral-platform choices eventually target the user, and that the only durable answer is self-hosting. The split comes down to who gets to decide what "mostly" means in a pull request at 2 a.m. on a Tuesday.

The first borderline repositories will land on Codeberg's moderators within weeks. The rule, by its own admission, does not tell them what to do.