The patterns picked up by the training don't seem to offer too much variation than a simple Markov chain. The author finds the generated texts similar to a conversation, because that's what they're looking for. But they look just as random as simply selecting the next random word that comes after the current word from the training set.