(a.) those graphs are a crime against data viz.

(b.) they practically demonstrate the point: while, yes, AI uses em-dashes, the entire corpus of em-dashes is still largely human, too, so using that as a sole signal is going to have a pretty high false positive rate.