No, they don't. They capture underlying patterns in training data and use them to generate text that fits the context.