Glossary
Hallucination (AI)
Definition
What is Hallucination (AI)?
In AI, a hallucination is a confident answer that is not supported by the model's input or by facts: an invented price, a policy that does not exist, a citation that was never written. It is the main risk in customer-facing AI and is controlled by grounding answers in your documents, restricting scope and routing uncertainty to a person.
Last updated:
Four controls that work
Ground answers in retrieved documents. Restrict the assistant to its domain and let it say 'I don't know'. Validate structured outputs against your systems. Log every conversation and review a sample weekly. Together these bring customer-facing error rates to a level below a busy human agent's.
Frequently asked questions
Can hallucinations be eliminated completely?
No, but they can be made rare and harmless: rare through grounding and scope, harmless by never letting the model act on uncertain information without a check.
Related terms
Workflows
Related services
Have a process like this? Tell us about it.
Book a call