What is a knowledge cutoff?
The knowledge cutoff is the point in time at which the text used to train a language model ends. Everything the model knows on its own, from tax rules to product names to who holds a public office, comes from documents written before that date. The model is typically released months after the cutoff and used for a year or more beyond it, so the gap between what it knows and the present keeps widening. A model can be excellent at analysis while unaware of a regulation that took effect after its cutoff.
Why the knowledge cutoff matters
Asked about a recent change, a model will often answer from the older state of the world with full confidence, which is a common source of hallucination in business use. A collections agency asking about current state licensing requirements, or a logistics firm asking about a new customs form, may get a confident answer that was correct when the model was trained. The fix is to supply current information in the prompt. Retrieval-augmented generation pulls in the company’s own up-to-date documents, and tool use lets an agent query live systems.
What to watch for
- Check the stated cutoff on the model card, and remember that coverage of the final months before it is usually thin.
- Never rely on a model’s built-in knowledge for anything that changes, including prices, laws, and personnel.
- Instruct the model to say when a question depends on information it may not have.