I mean, suppose the LLM bot is actually good at avoiding false positives/misunderstandings -- doesn't that simply remove one of the biggest weaknesses of old-fashioned keyword identification? I really just see this as a natural evolution of the technology and not some new, wild thing. It's just an incremental improvement.

What it absolutely does NOT do is replace the need for human judgement. You'll still need an appeals process and a person at the wheel to deal with errors and edge cases. But it's pretty easy to imagine an LLM bot doing at least as well a job as the average volunteer Reddit/Discord mod.

Of course, it's kind of a moot point. Running a full LLM bot, parsing every comment against some custom-design model, as your automoderator would be expensive. I really cannot see it happening routinely, at least not with current tech costs. Maybe in a few years the prices will have come down enough, but not right now.