AI models predict likely text. By default they do not look up facts in a database. When a plausible-sounding but false statement is the most likely continuation, the model may produce it confidently. This is called a hallucination.

Hallucinations are not random lying. They come from how the model works: it generates what fits the pattern, not what is verified. That is why a model can invent a citation, a quote, or a statistic that looks real.

Accuracy is also an efficiency issue. A wrong answer you act on, or have to re-ask and correct, wastes the tokens you already spent. Getting it right the first time is cheaper.