I've also noticed that very obviously LLM-generated comments are called out, and the community tends to agree, but those that have any plausible deniability are given far too much leniency, and people will over-index on the guidelines to give them the benefit of the doubt.
I don't think a captcha is the solution, as it'll degrade conversation by an OOM though.
But mostly in my experience it's otherwise perfectly normal users who at some point just decide to post something racist or bigoted, advocate violence (again, usually in political threads) or antisemitism or espouse some insane conspiracy theory nonsense. At that point I no longer care about anything else they might have to say.