ModelRefs / Hallucination — AI Glossary
Hallucination — AI Glossary
When an LLM produces fluent but factually incorrect, fabricated, or unsupported output. Types: intrinsic (contradicts source), extrinsic (unverifiable claim).
Overview
Hallucinations stem from the model interpolating plausible tokens when grounding is absent. Types: intrinsic (contradicts source), extrinsic (unverifiable claim). Mitigations: RAG, citation requirements, constrained decoding, calibrated abstention, and production evals.
Reference details
| Topic | evaluation |
|---|---|
| Last reviewed | 2026-06-24 |
Related terms
Example: Two failures, two different fixes
A source document says revenue rose 4% to £2.1m. Intrinsic hallucination: the summary says revenue fell 4% — it contradicts the source, and a source-grounded entailment check catches it because both halves are in front of you. Extrinsic hallucination: the summary adds “driven by the new Berlin office” — nothing in the source says this, and it may even be true in the world. No check against the source can refute it; it can only observe that the claim is unsupported. Requiring a citation for every claim is what catches the second, and the second is the one that damages trust, because it reads as informed.
Commonly confused with
Not every wrong answer is a hallucination. A model repeating an error present in its retrieved context is a retrieval failure; a model that is out of date is a staleness problem; a model hedging on a true claim is miscalibration. Attributing all four to hallucination sends you to the wrong fix — usually more prompting, when the problem is upstream.
When to use it
Reach for it when:
- As an evaluation category, split into intrinsic and extrinsic and measured separately
- When specifying grounding requirements: which claims must carry a citation
- In incident review, to distinguish generation failures from retrieval failures
Reach for something else when:
- As a blanket label for any incorrect output, which hides the actual cause
- As something RAG eliminates — grounding reduces extrinsic invention, it does not end it
- In user-facing copy, where naming the concrete failure is clearer than the metaphor
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Hallucination — AI Glossary.
Frequently asked questions
What is Hallucination?
When an LLM produces fluent but factually incorrect, fabricated, or unsupported output.
What concepts are related to Hallucination?
Closely related concepts include rag, grounding, eval, guardrails.