Pinning Down What an LLM Hallucination Is: Types, Causes and Boundaries

What is a hallucination in a large language model? Give a definition precise enough to build on. ```hint Name the yardstick An output can only be "made up" relative to something. Decide what the model's answer is being checked against before you write the definition. ``` ### Clarifying Questions - Should the definition cover only statements that are false about the world, or also answers that contradict the documents or conversation the model was given? - Is the definition for a general chat assistant, or for a grounded product such as a document summarizer, where the source text is the standard of truth? - Do reasoning slips, such as a wrong calculation built on correct facts, count as hallucinations or as a separate failure? ### What a Strong Answer Covers - A precise definition built on the gap between fluent, confident output and what the evidence supports - The main types, with concrete examples such as invented citations or a summary that adds a figure the source never stated - The underlying causes in how models are trained and how they generate text - Where the boundary lies with related failures such as outdated knowledge, reasoning errors and requested fiction - Why the definition matters for measuring the problem and for user trust ### Follow-up Questions - The model states something that was true when its training data was collected but is no longer true. Under your definition, is that a hallucination? - Why can a model hallucinate even when the correct document is sitting in its context? - How would you measure the hallucination rate of a summarization feature compared with an open-ended assistant?