It's is RLHF that dominates the style of LLM produced text not the training corpus.

And RLHF tends towards rewarding text that first blush looks good. And for every one person (like me) who is tired of hearing "You're making a really sharp observation here..." There are 10 who will hammer that thumbs up button.

The end result is that the text produced by LLMs is far from representative of the original corpus, and it's not an "average" in the derisory sense people say.

But it's distinctly LLM and I can assure you I never saw emojis in job applications until people started using Chatgpt to right their personal statement.