Hallucination
When an AI states something false with total confidence, like an invented fact, quote, or link.
Ungrounded
- Model guesses from memory
- Sounds confident
- Can invent facts
Grounded
- Model reads your real data
- Cites its source
- Says "not sure" when needed
What it is
A hallucination is an AI output that sounds right but isn't: a citation that doesn't exist, a wrong price, a made-up feature. Models predict plausible text, and plausible isn't always true.
You reduce hallucinations by giving the model the real facts to work from (your product data, the actual document), asking it to say "I don't know" when unsure, and checking important outputs before they reach customers.
When something matters, design for verification rather than trust.
How to ask for it
“Sometimes the AI makes stuff up. Make it stop.”
“To reduce hallucinations in the support assistant, give it only our help-center articles as source material, require it to link the article it used, and have it reply "I'm not sure, let me get a human" when no article covers the question.”
"Make it stop" has no mechanism; naming hallucination leads to grounding, citations, and a fallback.
You've seen this in
- CChatbots inventing refund policies
- AAI-written articles citing studies that don't exist
- AMade-up URLs in AI answers