I already answered the first part your question at the end of my previous comment. As to the idea that they become more useful/beneficial than actual human members, I personally find actually interacting with other humans more fulfilling than interacting with a machine in a social context, regardless of how 'useful' they are.

If AI gets to the point where it's all pervasive, indistinguishable from real people, and we somehow cannot devise a way of weeding them out of online communities, I'd honestly probably just stop participating socially in public internet spaces, and instead restrict myself to private invite-only communities that do more intense vetting to ensure only humans are in that space, or participate in and help build up local mesh-networks, where people can actually meet up with who they talk to online.

That would probably mean a lot less time spent online overall, and instead going outside and finding connection with people in real life.

It is correct, it is very frequently used in much of the material LLMs were trained on, and the AI is correctly using it, but generally to excess. It uses it so often, that it has unfortunately become a signal that something might be AI written. That is just the reality at this point, and I suspect Em-dash usage by humans will begin to fall off, at least online to avoid that association (until the AI is trained on material that makes it not use em-dashes, at least).