Well you could in theory for lets say AI generated images, train a neural network model that could pick up on artifacts in an image that only seems to be present in AI art, as well as AI generated texts, seeing how common a certain sort of text structure appears or something like emotion or sentiment analysis where an AI generated text doesn’t do as good in terms of presenting genuine emotions.

Of course it’s not 100% there yet. But to call them bullshit is closing doors that are not fully realised