During training the model saw enormous amounts of text and adjusted billions of internal numbers (parameters) so its predictions matched real language. It isn't storing sentences to recite. It's storing statistical relationships between concepts, which it recombines to fit your specific request.
This is why the same question can get slightly different answers, and why phrasing your prompt well changes the output so much. You're steering a prediction engine, not querying a database of fixed facts.
Updated July 2026