> They say the models were trained on a bunch of books…
Yeah, it's where I learned to use em-dashes as well.
> In these contexts the dashes can and do seem out of place.
Hmmm… For sure I use em-dashes in HN comments. I am not sure that I mentally differentiate as to whether I am in one scenario or another. (But to be sure I am not likely to leave an Amazon review though — so perhaps those contexts you called out self-select.)
I use em dashes in my comments too but this is Hacker News. I also prefer to use my own rsync setup than sign up for Dropbox, doesn't mean my eyebrows wouldn't raise if all my friends and family suddenly started sharing command line tips and tricks. It's self selection like you say.
But my point about the article not being convincing is just this: I can share my anecdotal evidence, you can too, we all go in a circle and it gets us nowhere. What I was expecting when I clicked the link was some actual data on dash prevalence in casual writing such as YouTube comments and a conclusion based on that data. What I got was more "Well if you look at this very particular kind of writing then extrapolate that to cover all writing then my point is made."