Hallucination

Hallucination refers to generated content that is fluent but factually incorrect or unsupported by source material.

The model sounds confident while inventing details that are not true or not in the documents you gave it.

Hallucinations arise when generative models produce fluent continuations that are not entailed by training data, retrieved context, or world facts. Mitigation includes grounding, constrained decoding, abstention, and human review.

Hallucinations are a primary enterprise risk for customer-facing and knowledge workflows.

• Risk discussions in AI governance
• Design of citation and RAG systems
• Eval suites for factuality

• Naming the failure mode improves product design
• Drives investment in retrieval and verification

• “Hallucination rate” depends on task definition
• Not all errors are hallucinations (e.g., retrieval misses)

• Larger models do not eliminate hallucinations
• Temperature zero does not guarantee factuality

Can RAG stop all hallucinations?

No. It reduces unsupported answers when retrieval is relevant, but models can still misread or invent beyond sources.

• NIST AI RMF

• Surveys on hallucination in LLMs (arXiv literature)