A hallucination is an output from a language model that is linguistically convincing and factually wrong — an invented source, a non-existent quotation, a plausible but incorrect figure. The term misleads, because it suggests an exceptional state. In fact the phenomenon follows from how the model works and is not a defect.
Why it is inherent
A language model produces text by continuously forming probable continuations. It does not check whether a statement is true — it has no concept of truth, only of probability. A well-sounding wrong answer and a well-sounding right answer are structurally the same from the model’s perspective. That is exactly why confidence of wording says nothing about correctness of content, and exactly why hallucinations are dangerous: they look like good answers.
Where they occur most often
Certain situations raise the risk markedly. Questions about specific details — figures, dates, names, reference numbers, sources — are more prone than questions about relationships. Topics with thin data are more prone than well-documented ones. And queries that imply a particular answer invite confirmation: ask for the study that supports a thesis and you are more likely to be given a study, whether or not it exists.
What actually helps
The most effective countermeasure is to give the model the facts rather than extract them from it. That is exactly what RAG does: the relevant passages are retrieved and embedded in the query, and the model formulates on that basis. It shifts the source of error from inventing to selecting — an error remains possible, but it is checkable, because the source is named.
Citation as a checking mechanism
Requiring citations adds to this. When every statement must point to a specific passage, checking turns from research into a click. In regulated environments that is the decisive point: what makes an AI system deployable is not freedom from error but verifiability. An unevidenced answer is worthless there even when it happens to be right.
Why they do not disappear
Better models hallucinate less often, but not never — and the remaining cases are harder to spot, because they occur in an otherwise more reliable setting. Anyone relying on the problem vanishing with the next model generation builds a system without a checking mechanism and then no longer notices the error. The architecture must assume hallucinations occur.
Practical consequence
A production AI system needs three things: facts from a controlled source rather than from model memory, a citation for every statement, and a clearly named human checking step wherever the consequences of an error are material. Missing any one of them does not make the risk smaller, only invisible.