I actually find the "AI idioms" rather less grating than emoji-vomit. That said, I don't know why certain LLM output seems to be full of the latter; certainly no real human writing I've seen has that style, but perhaps it's a result of training on data that probably should've been done without.
Patrick Wardle, the guy behind Objective-See, had that style in the 2010s when I first started following his work. I actually liked it at the time.