A blanket ban is really the only sensible thing to do so that no time is wasted for both sides (contributors know upfront that there's no point trying to get an AI-generated PR accepted - so they won't waste time creating one, and project maintainers don't waste time reviewing what might be broken AI slop - even if some AI generated PRs would be acceptable from a quality point of view).
When there's a grey zone then there will be lots of pointless discussions like "why was this AI-generated PR accepted but not mine" etc etc...
Perhaps you misunderstood my comment. I'm not advocating for vibe-coded AI-generated PRs, and I do think that blanket banning them is pretty reasonable for the reasons you stated.
However, I don't think that banning all AI-generated code is reasonable. Having an LLM generate a couple of functions or a bit of boilerplate in an otherwise manually coded PR should not invalidate it from being accepted if it's helpful.