Zig's Anti-LLM Contribution Policy and the "Contributor Poker" Argument
Date: 2026-04-30 Source: Simon Willison covering Loris Cro / Zig Software Foundation Raw: raw/rss/2026-04-30-simon-willison-the-zig-projects-rationale-for-their-firm-anti-ai-contr.md
TL;DR
Zig has one of the strictest anti-LLM policies of any major OSS project: no LLMs for issues, PRs, or bug-tracker comments. Zig Software Foundation VP Loris Cro published the rationale: open source review work is an investment in people, not just code. PR review time is justified because each new contributor is a long-term asset. LLM-authored PRs break this calculus — the time the maintainer spends doesn't grow a contributor. He calls it "contributor poker" — you bet on the person, not the cards.
The most cutting line from the rationale: "if a PR was mostly written by an LLM, why should a project maintainer spend time reviewing and discussing that PR as opposed to firing up their own LLM to solve the same problem?"
The Bun project (Anthropic-acquired in Dec 2025) runs a fork of Zig and recently shipped a 4× compile speedup with parallel semantic analysis. They explicitly say they will not upstream because of Zig's policy. So the policy is now affecting which performance work flows back into the upstream tree.
Why It Matters
This is the clearest articulation of an anti-LLM open-source position to date — and crucially, it is grounded in a project economics argument, not an ethics argument. The argument has bite even if you fully accept LLM-generated code. The choice is structural: PRs are training data for human contributors, and LLM-authored PRs short-circuit that training loop. Projects optimizing for long-term contributor health have a coherent reason to ban them; projects optimizing for code throughput don't.
The Bun divergence is the deployment consequence. Anthropic's flagship JS-runtime acquisition can no longer contribute back to its language's mainline. If more major Zig users follow Bun, the policy creates a permanent fork structure between AI-heavy and AI-free Zig ecosystems.
Connection to Prior Wiki Knowledge
Companion to Building Pi (04-29) and Claude Code postmortem (04-23-24). Mario Zechner (Pi) and Loris Cro (Zig) are making complementary observations from opposite ends:
- Zechner: the user of an AI agent struggles with behavioral unpredictability; specialized harnesses help.
- Cro: the project receiving AI-authored output struggles because review time stops growing contributors.
Both diagnose the same underlying fact — AI-assisted code shifts the cost structure of software development — and propose different mitigations. Zechner builds a stable harness; Cro builds a wall.
Connects to "code quality is down everywhere" (Pragmatic Engineer, 04-29). The 30+ engineering teams Armin Ronacher surveyed reported "vibe slop" shipping to production. Zig's policy is the OSS-side response to the same observation: filter at the contribution boundary rather than fix downstream. Whether the wall holds as AI-assisted code becomes universal is the open question.
Worth Watching (90 days)
- Other major OSS projects adopting similar policies. Zig is currently the most prominent. If even one of {Linux subsystem, Postgres, CPython, LLVM} adopts language with similar bite, the pattern crosses from outlier to convention.
- Bun fork divergence depth. If Anthropic invests in Bun's Zig fork as a serious infrastructure tool, the fork stops being temporary. That's the test of whether anti-LLM policies actually push major industrial users away.
- Compromise positions. A middle ground — "LLM-assisted PRs accepted with full disclosure and human-attributable rewrite" — has not yet emerged as a stable convention. Some project will try it.