It has been trained on tons of existing text written by humans, and so a large enough neural network can learn all kinds of low-level and high-level patterns that exist in it, from groups of words all the way to abstract concepts. When given a new input it basically fills in what an response would look like based on the patterns it learned. Basically there it goes the other way, from abstract concepts back to individual words, which is similar to how a human would think of something and then formulate a sentence. But its concepts are based only on all the texts it that it has been trained with, and not on anything else. So it only ever reinforces existing memes or consensus.