It's not so much an LLM thing as it is a data/curation thing. Idk the specifics, but I do remember back when I used it, chatgpt occasionally presented me with multiple responses, and asked me to pick the one that I liked best. It's like how Facebook/tiktok/whatever curate your feed by examining how much time you spend eg looking at one post vs another. There's nothing inherent about LLMs that cause them to prioritize emdashes, thats all the result of extra massaging the operator has applied to the model.
This is also why I feel it's important to not focus on perceived "LLMisms" as a way to detect slop. It's extremely trivial to hide that, and the ones that don't are low hanging fruit.