As well as being syntactically correct, it also reflects the contextual patterns of the tokens used to train it and then prompt it, which is perhaps limited by both intelligence and knowledge in humans.