I wouldn’t say they have an undefined truth value. Their source of truth is their training data. The problem is that human text is not tightly coupled to the capital T truth.