What is training data?
The text (and other content) a model learns from — its source of both capability and bias, and why a "knowledge cutoff" exists.
Training data is the text and other content a model learns from during its development — the source of both its capabilities and its biases. A model like ChatGPT, Claude, Gemini or Copilot has no experience of the world beyond what was in this material: it learns patterns of language and knowledge from examples, and everything it can later do traces back to something present, in some form, in that collection.
Why the cutoff exists
Training data is gathered up to a certain point, and the field describes the resulting boundary as the model’s “knowledge cutoff”: the date after which it has no awareness of events, because nothing published later was part of what it learned from. A model asked about something that happened after its cutoff cannot know the answer from training alone — it can only reason from what came before, or say plainly that it does not know. This is why any assistant’s knowledge is always a snapshot, not a live feed, unless it is separately connected to current information.
Why bias is baked in
A model trained on human writing inherits human biases, errors and gaps, because it is learning statistical patterns from what people actually wrote rather than from a curated, neutral account of the world. NIST’s framing of AI risk treats this as a structural property of the technology, not a flaw unique to one system: the data shapes the model, so imbalances or mistakes in the data become tendencies in the output. Researchers are still working out how much a model can be said to “understand” what it has learned, as opposed to reproducing patterns convincingly, and that question remains open rather than settled.
Questions people ask
Why doesn’t AI know about recent events?
A language model learns from a fixed collection of text gathered up to a certain point, called its knowledge cutoff, and once training stops it has no awareness of anything that happened afterwards. Having a conversation with it doesn't retrain it, so its knowledge stays frozen at that date no matter how you phrase your questions. Some systems get around this by searching the web or being handed current documents at the moment you ask, which is how they can discuss recent events despite the underlying cutoff.