The word "hallucination", applied to language models, describes a specific behaviour: the model produces plausible but invented information, presenting it in the same confident tone as something correct. It is not a system fault in the traditional sense — it is intrinsic to how these models work.
Why it happens
A language model generates text by predicting the most likely next word from patterns learned during training. When it has no precise information to draw on, it can still produce an answer that is grammatically correct and stylistically convincing — simply not anchored to a verifiable fact.
Where it is most likely
Hallucinations are most frequent when the model is asked for a specific, verifiable item — an exact figure, a precise quotation, a particular fact — that is not explicitly present in the context supplied. They are rarer in comprehension, summarising or dialogue, where the model works on information genuinely present in the conversation.
How to protect yourself in practice
The most effective defence is not hoping the model never errs, but designing the system so that critical data — calculations, prices, verifiable facts — always comes from sources verified outside the model, not from its free generation. The model reasons and communicates; the numbers come from a deterministic system upstream.