Large language models often diverge from human interpretation of probability words, affecting trust in critical decisions.