Every large language model was trained on text gathered up to a certain point in time. That point is its knowledge cutoff. Ask about anything that happened after it and the model is, by default, flying blind — it will either admit it doesn't know or, worse, confidently make something up.
This is the single most common source of confusion for new users: an AI that writes beautifully about history can be hopeless on last week's news, not because it's "dumb" but because that week was never in its training.
Why it matters at your desk. For a researcher or lawyer, the cutoff is a citation hazard — a model may not know a statute was amended, a ruling was overturned, or a paper was retracted after its training ended. Two things bridge the gap: tools that search the live web, like Perplexity, and feeding the model current documents yourself via retrieval-augmented generation (what Claude Projects does when you upload files). Newer releases such as Opus 4.8 push the cutoff later, but there is always a cutoff.
What to watch for: a model rarely volunteers "that's past my cutoff." If recency matters, either give it the up-to-date source or use a tool that can look it up — and never assume it knows today's facts just because it sounds current.