Models describe a world that stopped
Training data ends somewhere, and everything after that point is either missing or inferred. The model does not experience this as a gap. It answers in the present tense because the present tense is what the training text used.
So a draft can be entirely true of the world eighteen months ago and wrong about today, without a single internal inconsistency to warn you. There is no tense marker, no footnote, no hesitation.
The practical consequence: recency is not something to check after the fact, it is something to build into the claim. Anything that can change needs the date it was true written beside it.