Rust-lang/rust is adopting an LLM policy

TL;DR Highlight

Rust allows LLMs for questions and review but strictly restricts generated code, requiring disclosure.

Who Should Read

Developers who submit LLM-assisted PRs or issues to open-source projects, or who must review them.

Core Mechanics

The policy covers rust-lang/rust only
- Five teams adopted it, and it is not an official Rust stance on LLMs.
- It affects reviewers and moderators, authors of LLM-generated PRs, people who find issues with LLMs, and people who quote LLM output.
The rule: help, don't create
- LLMs are fine for answering questions, analyzing, distilling, refining, checking, suggesting and reviewing.
- Creating uses are heavily restricted, and LLM output in public docs, PR descriptions and comments must be clearly marked.
LLM code PRs face a higher bar
- Only pre-arranged, non-critical, high-quality, well-tested and well-reviewed LLM code is allowed, with disclosure.
- LLM PRs must have tests, and soundness-critical changes are barred unless the author is a domain expert.
Three problems motivated the policy
- Polished PRs no longer prove effort or understanding.
- Easier code writing worsens the existing review bandwidth shortage.
- Copy-pasting review comments into an LLM and posting its reply wastes time and breaks trust.

Community Reactions

Opinion that new contributors don't know reviewers
- A commenter said new contributors must talk to a reviewer before an LLM PR, but they know none.
- They asked what the proper procedure is.
Praise for anti-harassment and no-policing rules
- Some liked the ban on harassing LLM users.
- One said they had received severe insults on forums.
- Another especially liked the rule not to police whether someone used an LLM.
Concern that disclosure invites witch hunts
- A commenter feared disclosure rules lead to suspicion based on appearance.
- They argued it could make communities gatekeep newcomers.
Doubts about enforceability
- One commenter said people can hide AI use by removing em-dashes.
- Another noted Rust already has AI code in its codebase.

How to Apply

Add an LLM-use field to your PR template
- Today, add a checkbox asking whether LLMs generated, analyzed or translated any part, and which.
- Put the burden on authors so reviewers needn't guess.
Write a create-vs-assist rule
- In CONTRIBUTING, allow questions, analysis, summaries and review, and require pre-agreement and tests for generated code.
- Require marking LLM output in public comments.
Add LLM PR criteria to the review checklist
- Check tests, prior agreement and whether it touches critical areas first.
- If any fails, close it citing the policy clause.
Require human-written review replies
- State in team rules that review replies must not be pasted LLM output.
- If LLM text is used, it must be quoted and marked.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version adapting the Rust LLM policy ideas into a CONTRIBUTING draft)

You are an open-source maintainer. Using the principles below, draft an 'LLM usage policy' section for our repo's CONTRIBUTING file.
- LLMs are fine for questions, analysis, summaries, refining, checking, suggestions and review. Creating new content is heavily restricted.
- LLM output in public docs, PR descriptions and comments must be clearly marked.
- LLM code PRs require prior agreement, tests, and no critical areas.
- Reviewers may close non-compliant PRs, and authors must disclose LLM use.
- Suspecting LLM use from style alone and harassing LLM users are prohibited.

Our repo info: {{repo description}}
Areas we designate as critical: {{critical areas}}

Next step: ask for matching PR template checkbox wording too.

Terminology

soundness-critical changeA change that could break the safety guarantees the language promises. Like work on a load-bearing pillar, only experts should do it.

monorepoA single repository holding many components. rust-lang/rust is an example.

MIRAn intermediate representation the Rust compiler uses while transforming code. Optimization PRs touching it can be risky.

shotgunningThrowing many PRs at reviewers at once. It only raises their decision cost.