What is a knowledge cutoff?

A knowledge cutoff is the last date covered by the material a model was trained on. It limits one thing only: what the model can recall unaided. Hand it a search result, an uploaded file or text you paste, and it reads that perfectly well. Later fine-tuning can add knowledge too, so a fine-tuned model may know things the base model never saw.

The material used to build a model gets collected up to a certain point and no further, and that boundary is the cutoff. Ask a model about an election held after it and there is nothing stored to draw on. The awkward part is that it does not experience the gap as a gap, so instead of stopping it may assemble a confident answer from what it did read. Why not just keep training it as events happen? Because training is one enormous run costing a fortune, so what a model knows is a photograph rather than a live feed. This is exactly why web search and retrieval-augmented generation exist. Both work by fetching current material and placing it in front of the model at the moment you ask, which sidesteps the cutoff entirely. A model with search enabled can discuss this morning's news. The same model without it is guessing from a year ago. The difference is what it was handed, not what it is. Models are also unreliable about their own cutoff date, since that date is just another fact somebody had to write into the training material.

Someone back from a year at sea. Whatever happened while they were gone, they were not there for any of it. Hand them today's paper and they will discuss it as well as anyone, because there is nothing wrong with their reasoning. The problem only appears when nobody hands them the paper and they answer anyway, from how things looked when they left.