AI-Generated PR Spam Is Open Source's New Email Problem

When OpenClaw became GitHub’s fastest-growing repo, its pull request volume exploded from two a week to 3,400, and its merge rate collapsed from 48% to under 10%. Most of the flood was low-effort output from contributors’ own AI coding agents — one person filed 106 PRs in a day, a few seconds apart. Greptile, which runs AI review on the repo, argues this mirrors early-2000s email spam: once the cost of sending drops to near zero on a trusted platform, volume and abuse follow. The emerging fixes look familiar too — blocklists, confidence filters, and reputation scoring. OpenClaw already merges first-time contributors at 8.2% versus 18.6% for repeat ones, and Mitchell Hashimoto’s Vouch system for Ghostty functions as a sender-reputation layer that gates who can contribute at all.

A subtler problem is convergence. Linus’s Law assumes many eyes bring diverse perspectives, but when everyone prompts the same models — Claude, Codex, Cursor — contributions start to look alike. On OpenClaw, ten-plus people independently submitted the same SearXNG search-provider feature, six fixed the same Brave Search locale bug, and five found the same timeout deadlock. The diversity of thought that makes open source work erodes when the underlying reasoning is outsourced to identical tools with identical prompts.

What survives review tells the real story: refactors merge at 35% while novel features merge at just 9%. Changes that demand deep understanding of the existing codebase outperform greenfield construction by roughly 4x, because an agent told to “build a memory system” can’t reverse-engineer the non-obvious architectural tradeoffs a human who studied the system can encode into a prompt. The thinking matters more than the typing. The takeaway is that open source can now build at unprecedented speed, but it will need stronger primitives for identity, reputation, and contribution validation to absorb the resulting noise.