A customer-facing assistant can give a confident answer that its evidence does not support. The result may look useful while damaging trust: Air Canada’s assistant failure illustrates the reputational cost. A hallucination is a plausible generated claim that is factually wrong or ungrounded in the available source material. Unlike a software error that announces itself, it can fail silently. section 1.3 section 1.3.2
The distinction between error types helps locate the failure:
- Factual hallucinations contradict reality; citation hallucinations invent sources or links that do not resolve. section 1.3.2
- Intrinsic hallucinations contradict retrieved text; extrinsic hallucinations add details absent from it. A claim can therefore sound reasonable without being supported by the supplied evidence. section 1.3.2
- Logical hallucinations contradict the answer’s own reasoning; temporal hallucinations confuse historical and current information. section 1.3.2
When most of an answer is right
A legal research answer cited a real case but attributed it to a fictional judge. The accurate surrounding details made the invented detail harder to notice. In an open-domain question-answering comparison, a model answering from its parameters alone gave “Vancouver” as the largest city in the Pacific Northwest; a retrieval-augmented system combined passages about the region and Seattle to answer “Seattle.” section 1.3.2 section 4.5.1
Retrieval-augmented generation addresses this risk by locating external information and using it to generate an answer. Retrieval is not itself verification: diagnose whether relevant evidence was found, whether the answer used it faithfully, and whether each consequential claim is supported. Larger models and longer prompts alone do not remove the problem; polished fabrication can be difficult to detect, while excessive context can obscure key facts. section 1.3.2 section 1.3.4 section 1.6




































