What it means
Training data is collected up to a point and then frozen. A model has no awareness of anything after that date: elections, product launches, price changes, its own successors.
The dangerous part is that the boundary is soft rather than sharp. Coverage thins out near the cutoff, because commentary about recent events accumulates for months afterward — so a model may have fragmentary, unrepresentative knowledge of the period just before its stated cutoff while appearing confident about it.
Models are also often wrong about their own cutoff, since that fact has to be told to them rather than known by them. And a cutoff is not a release date: models frequently ship many months after their data ends.
Web search and retrieval are the standard fixes, which is why most consumer assistants now search rather than answering purely from training.
Why it matters
This is the most common source of confidently wrong answers about anything current, and it is invisible in the answer itself — a model does not say "this may predate my training." It just answers.
In practice
For anything time-sensitive, supply the current information or use a tool with live search. Do not ask a model what its cutoff is and trust the reply — check the vendor documentation.