Hallucination
A hallucination is model output presented as fact that is not supported by the material supplied to the model, nor by anything else — a fluent, confident statement with no basis.
The behaviour follows from what a language model does. It produces text that is likely given what precedes it, and plausibility is not the same property as truth. A well-formed sentence naming a document section that does not exist is exactly as easy for the model to produce as one naming a section that does, because both are ordinary-looking text.
There is no signal in the output distinguishing the two. Hallucinated content is not hedged, malformed, or otherwise marked; it arrives in the same register as the rest of the answer. This is what makes the phenomenon expensive rather than merely annoying.
The word is a metaphor and a contested one. It implies perception gone wrong, where the mechanism involves no perception at all, and it has been criticised for suggesting a malfunction rather than a routine consequence of how the models work. Confabulation, fabrication, and ungrounded generation have all been proposed instead. None has displaced it.
In practice
In a retrieval-augmented system the term covers several failures with different causes, and separating them is the difference between fixing one and chasing all of them.
Unsupported claims despite adequate sources. The right passage was retrieved and the answer states something it does not contain. This is a generation failure.
Answers invented because nothing relevant was retrieved. The model received unrelated passages and answered anyway. The visible symptom is a fabricated answer; the cause is retrieval, and the resolution lies upstream. See retriever.
Fabricated citations. A reference, filename, section number, or URL that does not exist, or that exists and does not say what is claimed. Reference-shaped strings are highly patterned, which makes them easy to generate convincingly.
Plausible completion of a partial source. The retrieved passage contains part of the answer and the model supplies the rest from expectation — a procedure gaining a step, a list gaining a member. The hardest kind to notice, because most of the answer checks out.
Faithful reproduction of a wrong source. Commonly reported as a hallucination and not one. The answer is supported by the supplied text; the text is incorrect or outdated. Nothing in the generation step is at fault. See grounding.
Supplying source material narrows the opportunity for the first four without removing it. The model retains its parameters and its tendency to produce coherent text, and both operate alongside whatever was retrieved.
Commonly confused with
Factual error. An error is a claim that is wrong. A hallucination is a claim with no source at all, which may happen to be right. The categories overlap but neither contains the other, and a correct hallucination is still a system that stated something it had no basis for.
Ungroundedness. The precise technical description of most of what is called hallucination: output not entailed by the supplied context. It is the measurable property, and it applies without any claim about whether the statement is true in the world.
Retrieval failure. The upstream cause of a large share of hallucinated answers. When a system answers a question whose material was never retrieved, the generation step is behaving normally on bad input. Attributing this to hallucination misdirects the diagnosis.
Reasoning error. A model deriving a wrong conclusion from correct supplied facts has made an inferential mistake rather than invented content. Both produce a wrong answer; only one involves unsupported material.
Refusal or abstention. Declining to answer is the opposite behaviour and is often penalised as unhelpfulness, which is part of why the failure persists — a model that answers anyway looks more useful than one that says the corpus does not cover the question.
Usage notes
The word is used for any wrong output, which drains it of meaning. Applied to reasoning errors, formatting mistakes, stale sources, and unsupported invention alike, it stops identifying a cause. Naming the specific failure is more useful, which is why evaluation writing prefers “unsupported” or “not entailed by the context.”
It is contested as terminology, not merely as a metaphor. The objection is that it frames a normal property of the mechanism as a fault, and that it implies an intent to report the world accurately. The word remains standard in both research and product writing despite the objection, and both usages should be expected.
Retrieval reduces the rate; it does not change the kind. A grounded system produces the same behaviour less often. Claims that a retrieval architecture eliminates hallucination describe an ambition, not a mechanism.
“Hallucination” also appears in image and speech contexts for invented detail in generated or reconstructed output — plausible texture that was never in the input. The same metaphor, a different subsystem.