LLVM contributors challenge ban on AI-generated communication
Critics say the project council adopted the rule without real consensus, while defenders call it an incremental response to maintainer burnout.
LLVM's project council has treated a ban on AI-generated communication as adopted, and a group of contributors wants that decision unwound before it hardens into written policy.
Joe Nash asked the project to revert the rule, arguing it is premature and already being applied before its scope is clear. In one case a reviewer told a contributor they were no longer allowed to paste model output, even though a human had approved every message sent. Nash said that example does not obviously violate the policy as written, and that LLVM cannot sensibly police AI-written comments without also deciding how it will review, prioritize, and rate-limit the flood of pull requests those tools make cheap.
Mehdi Amini objected more sharply to the process. He said he had raised strong opposition and that the council rolled over it rather than refining the text. "It is unacceptable to me that people (like me) who expressed a strong opposition and presented argument are just ignored and rolled over here." Another contributor noted that the specifics have not yet been committed to documentation, so there is nothing formal to revert.
Defenders framed the ban as a narrow response to a real cost shift, not a comprehensive AI policy. David Blaikie said waiting for a holistic answer would leave maintainers exposed. The rule, he wrote, targets people who act as a proxy for a model instead of answering a reviewer themselves: that is rude, and it produces what he called a social denial of service, in which cheap contributions impose expensive work on a small set of volunteers. A council member, writing in a personal capacity, said the November meeting found rough consensus after about a year of discussion and signals of urgency from the developer meeting. Rough consensus, that member argued, should not be defined tightly enough to be gamed, and conditioning any rule on solving every LLM problem would freeze the project. Amendments, a replacement communication policy, or a separate proposal were the suggested next steps, including an idea that model-generated material be visible only to people who opt in.
Nash accepted that maintainers need relief now, but proposed a different lever: state that reviewers owe no response time and may ignore submissions they do not want to engage, alongside rate limits and model-assisted triage. The council member rejected that framing. The harm is not only volume against a fixed pool of happy reviewers. Burnout is rising from dehumanized comments, strained attention, and the sense that careful work is being devalued. Explicitly blessing ignored pull requests, that reply said, would erase the maintainer's job, which is to see that wanted changes get timely review. Letting unwanted traffic hit the inbox was compared to a poor spam strategy.
Others questioned whether a how-it-was-written rule can be applied fairly. One contributor warned that detection tools become a cat-and-mouse game and can penalize non-native English or unusual style, and asked whether maintainers may still accept a change they merely suspect was drafted by a model. That contributor also asked for a BOLT exception, on the view that generated reproducers can be turned into real upstream tests when a human understands and owns them. Another said comments clearly labeled as model output should not count as a violation, and compared an imperfectly enforced ban to a speed limit: the limit can still be worth having if existing processes deal with clear breaches.