Hallucinations are the main reason AI output should not flow straight into customer communications, financial records or legal documents without controls. A model might cite a policy clause that does not exist or state a delivery date it never saw.
The risk can be reduced, though not eliminated. Grounding answers in approved sources with retrieval-augmented generation, requiring citations, instructing the model to say when it does not know and testing with evals all help. For example, a knowledge assistant can be designed to answer only from linked documents and to decline when nothing relevant is found. The misconception is that a newer model solves the problem. Controls still matter.