On July 23, 2026, Codeberg — the Berlin-based nonprofit Git hosting platform — published a new Terms of Use policy banning repositories whose content is primarily written by generative AI. The policy passed a member vote with 358 in favour, 144 against, and 14 abstentions — roughly 71% support at approximately 50% member turnout. The Terms of Use change was committed the same day.
This makes Codeberg the first major public code forge to adopt a hard prohibition on AI-generated repositories — a direct contrast to GitHub, which has embedded AI generation into its core product via Copilot.
What the Policy Actually Says
Section 7 of the updated Terms now reads: "You must not share projects that mostly consist of code written by 'generative AI'-tools (including services such as Claude, OpenAI Codex)."
The ban targets:
- Projects created or maintained primarily by autonomous LLM agents
- Projects whose CI/CD and storage consumption is disproportionate to the human effort behind them
- LLM-focused ecosystem tooling
It explicitly does not cover: established communities whose projects predate the LLM era, small experiments with low resource usage, or individual AI-assisted contributions within a human-maintained codebase. Importantly, human review of AI output creates no exemption — if the code is "mostly" AI-generated, it is still prohibited.
Passed alongside the AI ban: a commitment that Codeberg will never train AI models on hosted code or user data. A cryptocurrency project ban was also approved in the same vote.
The Reasoning
Codeberg's official blog post gives four connected reasons. Infrastructure first: SSD hardware costs at the nonprofit have risen from roughly €700 to €3,700 per unit, partly driven by AI datacenter demand, and repositories spun up by autonomous agents with heavy CI/CD pipelines are unsustainable for a donation-funded organization. Second, copyright: AI-generated code has unresolved legal status, creating liability risk for the platform. Third, security: heavily AI-generated projects may ship insecure code without meaningful human review. Fourth, and most explicitly: "LLMs endanger the free/libre software ecosystem as a whole."
How It Compares to GitHub and GitLab
GitHub treats AI generation as a competitive feature — Copilot is built into the product and public repository data is used for training by default. GitLab's Duo AI avoids storing user data outside approved collections but places no restrictions on AI-generated repositories. Codeberg now sits at the opposite end: no AI training on its data, and no predominantly AI-generated codebases on its platform.
Armin Ronacher — the Flask and Jinja creator — published a detailed critique arguing that "mostly" is undefined and unenforceable, that AI detection at scale is technically unreliable, and that a narrower policy targeting concrete harms (autonomous repository spam, abusive resource consumption) would be more defensible than policing authorship. It is a reasonable concern. Enforcement is currently expected to be complaint-driven, not proactive scanning.
What It Means
No mass deletion of existing repositories is planned. For the broader open-source hosting landscape, this vote sets a clear precedent: there is now a significant public forge that has decided the right answer to AI-generated code is prohibition, not tooling. Whether that position holds under enforcement pressure — and what "mostly" ends up meaning in practice — will be worth watching.