Rust Adopts LLM Usage Rules for rust-lang/rust Contributors

Five teams within the Rust project have adopted a policy defining how large language models (LLMs) may be used when contributing to the rust-lang/rust monorepo. The document's author, Jynn Nelson, stresses that this is not the project's official stance on AI, and the rules don't apply everywhere — they concern only those who review and moderate PRs, submit generated code, use LLMs to find bugs, or quote models in discussions. Everyone else needs to change nothing about how they work.

Why the rules were needed

Until now the project had, in the author's words, a “wild west”: dozens of PRs from language models, no requirement to disclose their origin, newcomers trying to push risky MIR optimizations as their very first commit. Moderators had nothing to point to, and new contributors didn't understand why their work was being closed.

Nelson identifies three problems that LLMs have brought. First, a polished, well-tested PR is no longer evidence of effort invested or understanding — and in the case of autonomous agents there may be no human “on the other end” at all. Second, the ease of writing code aggravates a long-standing woe: there are far more people eager to write code than to review it (at the time of publication, 1,281 PRs were open). Review is first and foremost about deciding whether such an approach is even needed, and the code itself is the least important part of it. Third, mechanically copying a model's answers in response to a reviewer's comments is a waste of time and an erosion of trust: “If we wanted an LLM's opinion, we'd ask it ourselves.”

What the policy says

The gist can be put in a single phrase: models may be used to answer questions, analyze, clarify, verify, suggest, and review — but not to create. Generated text is prohibited in documentation, PR descriptions, and comments without an explicit label. Disclosure is mandatory for machine translation, “trivial” edits, bug hunting, and reviewing others' code with the help of AI. At the same time, messages in one's native language are welcome — translating into English is not required.

Code from LLMs faces stricter requirements than human code: tests are mandatory without exception, and changes critical to correctness (soundness) are prohibited unless the author is themselves an expert in the area. Some of the rules are knowingly unenforceable — and that, the author says, is not a bug: the goal is not to catch every violation but to draw a clear line that lets moderators judge by actions rather than intentions. It is spelled out separately that harassing people for using AI is not allowed, and that writing style is not considered proof.

Nelson acknowledges that not all the rules are good, but written rules are better than unwritten ones. Rust runs on consensus, there is no unified view of AI in the community and probably never will be, so it is impossible to ban or permit LLMs “from the top down.” The leadership council is already considering the creation of a separate subteam for AI policy, and the policy itself is written so as to be easier to change in the future.

Source: Lobsters