Zig creator: LLM-assisted PRs leave a 'digital smell' maintainers can detect

Andrew Kelley, creator of the Zig programming language, pushed back on the assumption that maintainers can’t distinguish AI-generated contributions from human ones. He argues the failure modes diverge sharply: human mistakes follow recognizable patterns of misunderstanding or oversight, while LLM hallucinations carry signatures that are obvious to reviewers who don’t use the tools themselves. Contributors steeped in agentic coding workflows, he says, develop a stylistic tell they can’t perceive in their own output.

The analogy he reaches for is smoking — non-smokers detect it instantly on someone walking into a room, even when the smoker is oblivious. Kelley isn’t condemning the practice broadly, but he’s drawing a line at his own project: don’t bring it into Zig’s PR queue. The quote captures a hardening stance among some open-source maintainers who view undisclosed LLM-assisted contributions as a form of project-hygiene violation rather than a neutral productivity gain.