OpenAI recently launched an algorithm that purports to classify human- and AI-generated text, and for whatever reason they also decided to release actual precision/recall numbers which show how godawful it is. as OP points out, 30% of human-written text falls into the top 2 categories (as opposed to 54% of AI-generated text) but the fun thing is that in practice the numbers will be way worse than they even appear here for any application you're going to use it for, thanks to the magic of Bayes' Law.
if you have 1,000 students, 10 of whom commit AI-based plagiarism, 5 or 6 of them will be detected, compared to 297 honest students who are wrongfully accused (nearly 50 false positives for every true positive); if you decide to be more conservative and only use the top category as an indication of probable guilt, the numbers improve to a mere 3 plagiarists detected, against 89 wrongfully accused (it's correct almost 4% of the time!)
it was extremely clear from the jump that "starting up a secondary business detecting the pollutants we are pumping, as our primary business, into earth's information streams" was the goal here. and like most tech products it can't even do what it claims.
