It doesn't try that. It just guesses the next word, word for word, based on probabilities, which themselves are based on the training data. And after that baseline instruction, overall, the probability for phrasings like that is hightened. Though it still vanishes most of the times, when it's...