If we're talking about making an AI which generates text, then I agree. But I'm talking about an AI which classifies text (that I expect will be written by humans).

> isn't based around the word structure

Well, yes and no. REALLY stupid text has an identifiable structure. Midwit text looks smarter than it is. What I'm looking for is how to make the model deep enough to identify quality fedi banter.

Of course midwit diarrhea is a moving target w/ Goodhart's law, especially if people are start training GPTs against my "quality posts" classifier...