Hallucination happens because language models generate text by predicting what comes next, not by checking facts against a verified source. The model can produce information that sounds correct and well-structured but is simply wrong — an invented statistic, a non-existent policy clause, a made-up vendor name.
Hallucinated outputs often read as confidently as correct ones. There is no automatic flag that tells the reader something is fabricated. That combination of confidence and inaccuracy is what makes hallucination a real operational risk. Grounding AI outputs in real, retrieved data is the primary way to reduce it.
