Glossary
What is AI hallucination?
When an AI model states something that sounds right but isn't supported by any source, like an invented return policy, feature or price.
Language models write fluent text even when they don't have the facts, so they can invent a return policy, a feature or a price. In customer support that's worse than no answer.
The usual defenses: answer only from retrieved content, show the sources, and reply "I don't know" with a handoff when the content doesn't cover the question. Watching which answers get a thumbs down and reading their sources catches the rest.
See it in practice in Office Assist.
Where you'll use it
- Answers with sourcesEvery answer shows the pages it came from. When your content doesn't cover a question, the assistant says so and offers your team instead of making something up.
- Playground and answer inspectionChat with your assistant as any kind of customer and see exactly how each answer was made, then fix a bad one in a click.
Related terms
- GroundingTying an AI answer to specific source material, so every claim in it can be traced back to a page, article or document you provided.
- Retrieval-augmented generationRAG: a method where an AI model first looks up the most relevant passages in your own content, then writes its answer only from those passages.
- AI guardrailsRules and checks that keep an AI support agent on topic, within your policies and away from actions it shouldn't take on its own.
Try Office Assist with your whole team
7-day trial with every feature. Cancel from the billing page whenever you like.