"If AI labs wanted to go beyond that, they’d have to go and buy older books, which would probably have more em-dashes."

Actually, they wouldn't have to go and buy these old books: The texts are already available copyright free, due to legislation stating that copyright expires 70 years after the author's death (any book published in the USA before 1923 is also reproducible without adherence to copyright laws), making the full texts of old books much easier to find on the internet!