Bot prevention is a very difficult constant game of cat and mouse, and a lot of bot operators have become very skilled at determining the hidden metrics used by platforms to bless accounts; that's their job, after all. I've become a big fan of lobste.rs' invitation tree approach, where the reputation of new accounts rides on the reputation of older accounts, and risks consequence up the chain. It also creates a very useful graph of account origin, allowing for scorched earth approaches to moderation that would otherwise require a serious (and often one-off) machine learning approach to connect accounts.
Additionally, lots of Chinese and Russian keyboard tools use the em dash as well, when they're switching to the alternative (en-US) layout overlay.
There's also the Chinese idiom symbol in UTF8 which gets used as a dot by those users a lot, so that could be a nice indicator for legit human users.
edit: lol @ downvotes. Must have hit a vulnerable spot, huh?
That’s why the analysis was performed over time. All of those em dash sources you mentioned were present before LLM written content became popular.
I'm not trying to negate the fact. I'm just pointing out that a correlation without another indicator is not evidence enough that someone is a bot user, especially in the golden age of rebranded DDoS botnets as residential proxy services that everyone seems to start using since ~Q4 2024.