In plain terms
A model is like an expert who read everything up to a certain day and has been without newspapers since. Ask about anything before that day and they can answer from memory. Ask about last month and they either say they do not know or, worse, answer as if the world stopped on that day.
Why it matters
Models stay in service for a year or more after their cutoff, so the gap is always there. Prices, regulations, product versions, office-holders and recent events can all be out of date in a model's answer. Any application that depends on current facts needs a live source: search, a database or documents placed in the context.
Example
A user asks an assistant without web access for the current corporate tax rate. It states the rate that applied when its training data was collected. The rate changed eight months ago. Nothing in the answer signals that it might be stale.
Most often confused with
Knowledge Cutoff vs. Release date
The two dates are months apart: training, testing and safety work take place between them. A model released this spring may know nothing after last summer. When a model's recency matters, check its cutoff. The launch date tells you little.
Under the hood
The cutoff is a property of the training corpus, and it is softer than a single date suggests: the last months before it are thinly represented, because the web has not yet finished writing about them, so knowledge of that period is weaker. Models are often unsure of their own cutoff and of today's date; supplying the current date in the system prompt is standard. Remedies are all at inference time: web search tools, RAG over current documents, and instructions to prefer supplied sources over recalled ones. Fine-tuning can add newer knowledge but is a poor way to keep facts current.