Well I mean it’s a statistics machine with a seed thrown in to get different results on different runs. So really, it models the structure of language, but not the meaning. Kinda useless.

It's because, for the most part, it doesn't actually have access to the text itself. Before the data gets to the "thinking" part of the network, the words and letters have been stripped out and replaced with vectors. The vectors capture a lot of aspects of the meaning of words, but not much of their actual text structure.