knowledge staleness

A model's knowledge freezes on the day its training data was collected — its cutoff. Everything after that point is, to the model, the future it cannot see. Ask who holds an office that changed last month, the price of something that moved yesterday, or the result of a game played this morning, and a frozen model can only answer from the world as it was, often without realizing the world has moved on.

Two traps make this worse. The model rarely volunteers how old its knowledge is, so a confidently outdated answer reads exactly like a current one. And it can quietly mix a stale fact with present-tense phrasing, telling you who is the current leader using information that has since expired. The remedies are to give the model fresh information at question time — through search or retrieval — and to ask it directly what its cutoff is rather than assuming it is up to date.