[thomasahle]

That's pretty much how most of these methods work. It just doesn't work very well because good models have a reasonable probability of generating lots of different texts. So you don't get very different numbers on AI and Human generated texts. After all the models are trained to learn the probability distribution of exactly Human text.