I just don't know how I'd be sure a comment isn't written by an LLM.

Maybe fun little bike shed:

Are LLMs not just fancy Markov chains? They are next token predictors which have some hidden internal state that output probability distributions which lead to further states.

>A good old markov chain can simulate the average reddit thread

Someone did this on Reddit 9 years ago. It was remarkably good.

https://old.reddit.com/r/SubredditSimulator/comments/3g9ioz/

Bots used to just take other popular comments and repost them either in whole copies of threads, as in the example here, or taking a top comment in a new thread and reposting it elsewhere in the same thread. Now they're using LLMs to rephrase comments to try to avoid detection (though they often come across sounding a bit off so they're sometimes easy to spot).