It's utterly uncommon in the kind of casual writing for which people are using AI, that's why it got noticed. Social media posts, blogs, ...
AI almost certainly picked it up mainly from typeset documents, like PDF papers.
It's also possible that some models have a tokenizing rule for recognizing faked-out em-dashes made of hyphens and turning them into real em-dash tokens.
On my own (long abandoned) blog, about 20% of (public) posts seem to contain an em dash: https://shreevatsa.wordpress.com/?s=%E2%80%94 (going by 4 pages of search results for the em dash vs 21 pages in total).
AI almost certainly picked it up mainly from typeset documents, like PDF papers.
It's also possible that some models have a tokenizing rule for recognizing faked-out em-dashes made of hyphens and turning them into real em-dash tokens.