Survey of Hallucination in Natural Language Generation

Ziwei Ji and colleagues survey Hallucination across language-generation tasks. Their definitions distinguish contradiction of a source from additions it cannot verify, and faithfulness to input from factual accuracy (§2).^[1]

Why read it here?

Use its terminology to ask what an answer is grounded in. The Air Canada case illustrates the stakes of conflicting advice; the survey does not investigate that chatbot.

Historical context

First submitted February 8, 2022. The reviewed v7 revision is dated July 14, 2024; its first-page footnote marks the LLM section as updated in January 2024. Original metadata and selected definitions were checked, including rendered page 4. Versions, underlying studies and mitigation effectiveness were not independently audited.^[1]