How outdated information hides in LLM token generation probabilities