You're overthinking it. LLMs exploded the prevalence of em-dashes. That doesn't mean you should assume any instance of an em-dash means LLM content, but it's a reasonable heuristic at the moment.
I dunno, I feel like the base rate fallacy [0] could easily become a factor... Especially if we don't even have an idea what the false-positive or false-negative rates are yet, let alone true prevalence.
> That doesn't mean you should assume any instance of an em-dash means LLM content
No, it doesn't. But people are putting that out there, people are getting accused of using AI because they know how to use em dashes properly, and this is dumb.
