LLMs calculate a probability distribution for the relative chances of the next token, then select a token randomly based on those weightings.