The word is borrowed loosely from psychology. The model is not seeing things; it is filling a gap in what it knows with text that fits the pattern of a good answer. Hallucinations are most likely when the question needs a specific fact the model was never given.
Why it matters
In customer email a hallucination is a promise your business did not make. “Yes, we offer free returns within 60 days” is a pleasant sentence and, if your policy says 30, a problem you now have to walk back. Because the text reads well, the error is easy to miss when you skim.
Three things reduce the risk: grounding the answer in your own sources, showing which source each fact came from, and keeping a person between the draft and the customer (human in the loop).
An example
A customer asks for a discount code for students. No such discount exists, but a model without access to your pricing writes a friendly reply with a code that looks real.
How to catch this before sending: checking an AI-written answer.