They guess the next token based on probability which is derived from training data.