Overusing dashes is one thing, but I think the influence goes both ways. I didn’t use em dashes much in the past, but after seeing more examples—whether from ChatGPT or elsewhere—I’ve started incorporating them more often. The same goes for semicolons. (I used to rely more on parentheses, too.) For me, it’s a bit like picking up a new word: I never had a fixed or strongly defined writing style to begin with, so my style has shifted quite a bit over the years—whether due to LLMs or not. It feels more like an organic evolution than anything deliberate.

While em dashes might be extremely prevalent in AI generated text, if you plug in your priors, P(AI|em) is likely much smaller than you think it is.

After updating my false alarm rate, P(AI|em) now = 0.016. Before it was 0.09. People are simply over-updating without actually understanding how conditional probabilities work.