Why does it happen?
A language model predicts the most likely next word; it isn't looking for truth. If it doesn't have the facts, it still generates a fluent, confident answer, sometimes entirely made up. The model doesn't "know" that it doesn't know.
A hallucination isn't a lie; the model guesses convincingly.
Why is it a risk for businesses?
An AI that answers without your data can give a customer a wrong figure, a non-existent rule or an incorrect procedure, and it sounds just as credible as a correct answer. False confidence is more dangerous than "I don't know".
How do you reduce hallucinations?
- RAG: give the model your real sources at the moment of the question, so it relies on them.
- Source references: have the assistant point to the page or document the answer came from.
- "I don't know": let the model admit when info is missing instead of guessing.
- Human review for sensitive answers.
Related terms
- RAG: the technique that helps prevent hallucinations.
- Vector store: supplies the right sources to the model.
- Embeddings: make retrieving those sources possible.
