Hallucination

A Large language model (LLM) generates plausible language, not verified fact. When it does not actually know an answer it does not say so; it fills the gap with something that sounds right. That is a hallucination: invented hours, a made-up policy, a price you do not charge, or a dish you do not serve, delivered in exactly the same confident tone as everything it gets correct.

This is the core risk in letting AI speak for you. A phone or chat agent that hallucinates your hours, promises a reservation you cannot honor, or quotes an allergen answer wrong damages trust and creates real service failures, sometimes with safety stakes. It is why the honest advice is to point AI at lower-stakes tasks or keep a human in the loop, and why operators who use AI still skew toward lower-risk work (19%/15% marketing, 10% admin, 6% orders) rather than handing it the guest relationship unsupervised. AI companies fight the same problem partly by paying people to rate and fact-check model answers, and AITraining.jobs' guide to AI rater jobs, which we publish, explains what that work involves and what it pays.

Hallucination is the whole reason grounding matters. Connecting the model to your real, current data with Retrieval-augmented generation (RAG), writing a clear Prompt that supplies the facts, and choosing tools that cite their source all pull answers back toward truth. When you evaluate any AI vendor, the single most useful question is: where does its answer come from, and can it show you. See AI phone answering.