Examples
Worked examples
- Is an instance
An LLM citing a paper with a real-looking DOI that resolves to nothing
- Is an instance
A chatbot attributing a plausible but invented quote to a named scholar
Counter-examples
Looks similar, but isn't
- Not an instance
An LLM giving a wrong answer because the user's prompt was ambiguous (this is misunderstanding, not hallucination in the technical sense)
Editorial commentary
A hallucination is an output from a generative AI system that is presented confidently and fluently but is factually incorrect, fabricated, or unsupported by the input data or any verifiable source — invented citations, non-existent authors, false statistics, misattributed quotations, or plausible-sounding but wrong technical claims.
Hallucination is not a synonym for “wrong”
An ordinary factual error has a traceable cause: a misread source, an outdated figure, a miscalculation. A hallucination has no such traceable cause in the world — the model is not reporting a mistaken belief about something real, it is generating text that is statistically plausible given its training distribution and the prompt, with no underlying representation of “this claim is true” versus “this claim is invented.” The more useful comparison is to confabulation in the clinical-neuropsychology sense: a fluent, internally consistent, confidently delivered account that isn’t grounded in the events it purports to describe. A generative model producing a hallucinated citation isn’t misremembering a real paper; it is producing a citation-shaped string (plausible author names, a plausible journal, a plausible year) because that is the statistically likely completion, not because any such paper exists.
Why retrieval grounding reduces but does not eliminate it
Retrieval-augmented generation — having the model draw from a provided set of real documents rather than generating purely from its parametric training — measurably reduces hallucination rates, because the model has real source text to condition on rather than only its training-time statistical patterns. It does not eliminate the problem for three reasons: the model can still misattribute a real claim to the wrong retrieved source, still synthesise a plausible-sounding but unsupported inference from real retrieved passages, and still hallucinate when the retrieval step itself returns weak, irrelevant, or no matching source and the model fills the gap anyway rather than declining to answer. Grounding narrows the space in which fabrication can occur; it does not remove the model’s underlying lack of a true/false representation of its own output.
What it means for citation integrity
Because a hallucinated citation is typically well-formed — a real-looking author, journal, and year — it will not visibly announce itself as false, which is exactly what makes it a research-integrity risk rather than a nuisance. Any AI-generated factual content used in scholarly work (citations, statistics, quoted figures, attributions) must be independently verified by the human author against the actual source before it appears in a manuscript; the human author remains responsible for accuracy regardless of which tool produced the draft text (see author responsibility for AI use). This is the specific failure mode that AI research-search tools like Consensus, SciSpace, and Undermind, and AI data-analysis tools like Julius AI, are all built to reduce through source grounding and citation linking, and why each still recommends the user check the underlying source rather than trust the summary alone. See also AI output verification.
Sources
Ji et al., “Survey of Hallucination in Natural Language Generation,” ACM Computing Surveys (2023).
Also known as
Confabulation · AI fabrication
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Hallucination"
vocab-term-identifier="https://casrai.org/dictionary/term/hallucination" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/hallucination",
"name": "Hallucination",
"identifier": "https://casrai.org/dictionary/term/hallucination",
"description": "An output from a generative AI system that is presented confidently and fluently but is factually incorrect, fabricated, or unsupported by the input data or any verifiable source — including invented citations, non-existent authors, false statistics, and incorrect quotations.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/genai-disclosure#set",
"url": "https://casrai.org/dictionary/term/hallucination",
"sameAs": [
"Confabulation",
"AI fabrication"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"dateModified": "2026-08-22T14:39:10",
"inLanguage": "en"
}
