I bet a lot of people would get away with it, too. These generated texts are often just convincing enough to make you think they would make sense if you were less tired/distracted/whatever. A teacher in a rush or at the end of the day could easily be fooled (especially considering the quality of many human-generated term papers).
Now we need an equivalent of a "deepfake recognizer" for these sort of generated texts, to distinguish texts with some sort of coherent meaning from texts without.
As a bonus, such a model could be used in a word-processor to redline places where a human writer has written a lot without saying much.
