My top gripe with AI authorship is that it's very easy to create confidently sounding yet sub-par contributions with it, and that with very little effort. Applies just as much to PR descriptions as it applies to code. And when it happens the human in me finds the attempt annoying, in that one person meaning well yet being lazy and the other wanting to mean well yet needing to do extra work is in conflict, and as such not a sustainable daily routine.

An equilibrium emerging from the conflict appears to be to allow the reviewer to be lazy as well, categorizing contributions by perceivable AI-iness with little effort, for example by answering the questions "Is the PR title and description obviously human-authored?" or "Is the suggestion a reasonably minimal diff to achieve the stated goal?". And if the answer to any of these is "No", politely request an entirely new attempt on a mutually acceptable basis.

Long story short, yes we should try, but I also doubt that any guidelines can capture and prevent the disconnect, because one party is already lazy and has probably been convinced by confidently sounding assertions that what is being produced is of high quality. If necessary, a more absolute lever here is exactly what CountBleck mentions, in that AI as pursued nowadays is effectively an enormous pile of copyright infringement, so if we cannot be reasonably sure that a human has authored a contribution, we sadly need to reject it.

Given the widespread use of AI assistants & agents, PRs created with their help will inevitably arise. The question is what proportion is AI and what proportion is human involvement. One way or another, everyone uses AI to some extent, and many OSS projects starting from LLVM to Apache are trying to systematize and formalize this in some way, so that we can at least understand how to handle such PRs. As you and CountBleck noted, such PRs can contain many risks: copyright issues, lack of edge case coverages, the problem of reward hacking by LLMs, hallucinations etc. In short, the main idea is force to share the AI usage state and exactly how.

I'll add the relevant points to PULL_REQUEST_TEMPLATE.md in separate PR