LLVM RFC seeks ban on AI text in PRs and proposals

LLVM developer Nikita Popov has proposed tightening the project’s AI policy by forbidding AI-generated or substantially AI-assisted textual communication in pull request descriptions and RFCs, while leaving AI assistance for code unchanged.

The current policy is liberal as long as a human takes responsibility. Popov argues that experience now shows extractive use is common: many incoming PRs, especially from newer contributors, carry extremely verbose, heavily formatted descriptions full of routine testing detail and other noise. Those texts cost scarce reviewer time both to read and to rewrite into mergeable commit messages. He also contends that writing the summary oneself is a useful sanity check under the “human in the loop” model. Under the draft rule, any allowed AI-assisted prose elsewhere would have to be clearly labeled and collapsed; PR descriptions, which become commit messages, would contain none.

Reaction on the LLVM Discourse was mixed. Mehdi Amini called a total ban on AI wording in descriptions overly strict, saying carefully steered models can produce concise, high-quality summaries that beat what many authors write by hand, and warned that enforcement would devolve into arguments over “AI smell” instead of ordinary commit-message standards. Others preferred clearer anti-slop rules even if that forced some people back to writing a few paragraphs themselves. A further view held that the project should judge the quality of the text and the author’s understanding of it, not the tool that helped produce the wording, and should publish sharper guidance on good PR descriptions instead.

Several maintainers backed a harder line. Shafik Yaghmour and Erich Keane described burnout from large volumes of low-quality, LLM-driven PRs that stay extractive even after dozens of commits, with authors never graduating to doing more of the work themselves. The load is also complicating triage. The RFC does not change rules for code generation and still contemplates limited uses such as translation or proofreading, though participants noted the final wording will need to draw that line carefully.