Zig leaves GitHub for Codeberg — and bans AI contributions outright

Zig creator Andrew Kelley confirmed two decisions in an InfoQ/JetBrains interview: the Zig repository has moved to Codeberg (a code host run by a German nonprofit), and the project’s code of conduct now bans AI-generated contributions outright.

The facts

• The move’s trigger: GitHub Actions CI broke and stayed broken — “GitHub became completely unusable for us”; CI recovered after the migration. Kelley frames the choice as nonprofit stability: “nonprofits just commit to doing what they’re doing, and that stability is what I’m after.”
• The No-AI policy: ziglang.org’s code of conduct lists a Strict No LLM / No AI Policy — AI contributions are “garbage without exception,” negative value that adds review burden.
• The “contributor poker” metaphor: review time is a bet on a contributor’s growth trajectory; AI users are drive-by contributors not worth the stake.
• The background: Bun’s AI-assisted four-month rewrite of Zig into Rust drew Kelley’s criticism as a values clash, not a language verdict.
• Timeline: Zig began around 2016-17; Kelley went full-time in 2018.

Open-source governance picks a side

Rejecting both platform conventions (GitHub hosting and AI contributions) in one move is a first for a flagship project. Codeberg answers “who owns the infrastructure” — nonprofit, ad-free, no corporate politics; the No-AI policy answers “what is a contribution” — review time is the project’s scarcest resource.

The reference for other projects

Small, high-review-cost projects (compilers, crypto libraries, security tools) are the likeliest followers: AI-generated contributions have the highest false-positive rates exactly there. Large foundations (Linux, Apache) lean the other way — paid maintainers and machine triage absorb the flood. The middle will watch how Zig’s experiment ages.

Editorial take

Kelley’s two decisions are one judgment: project time is scarce, and both the platform and the contribution policy exist to protect it. The Bun episode proved AI can translate Zig efficiently — but efficiency was never what Kelley was optimizing; the question is whether a community can grow people who understand the system. Every maintainer buried in AI PRs should watch this experiment; next to arXiv’s rate-limit policy, infrastructure now has two answers to the AI flood — quotas and refusal.

The Codeberg trade-off, stated plainly

Codeberg gives Zig nonprofit governance, EU jurisdiction and no ads — at the cost of GitHub’s network effects: discoverability, Actions marketplace, and the default assumption that every project lives there. Kelley is betting Zig’s contributor pipeline is mature enough not to need GitHub’s top-of-funnel. The CI outage that triggered the move is also a data point about GitHub’s own reliability under AI-scale load — thousands of repos running the same runners.

What the No-AI policy actually bans

The policy governs contributions, not usage: developers may use AI privately, but what lands in the repo must be human-authored and human-understood. Enforcement is maintainer discretion — reject-on-sight plus a code-of-conduct clause. The interesting test is edge cases: AI-refactored patches, AI-explained bug reports, AI-translated docs. Zig drew the line at “if an LLM wrote it, it is not a contribution”; most projects will draw it somewhere fuzzier.

The practical question for contributors is where the line sits in review practice: a patch drafted with AI assistance, understood and defended by its author, likely passes; a raw LLM dump does not. Zig has chosen to make that judgment explicit and enforceable rather than aspirational — the part most projects get wrong is writing the value down. Codeberg’s nonprofit funding model is the quiet hero of the story.

Zig’s versioning discipline — 0.x for a decade — suggests a project comfortable saying no; this is the same instinct applied to platforms and contributors.

The Codeberg move also de-risks Zig from single-vendor dependency: GitHub outages, pricing changes and Copilot-mandated terms are all single points of failure that a nonprofit forge disperses. Projects with long horizons — languages especially — have different hosting economics than apps, and Kelley is pricing that in a decade ahead. The InfoQ interview is the primary source; Kelley’s patience with the questions is itself a governance artifact.