Hallucination is the term for output that reads perfectly and is not true.
It follows directly from the objective: the model was trained to produce likely text, not verified text.
A citation with a real-sounding author, a plausible journal and an invented page range is highly likely text.
So the model produces it with exactly the same fluency it uses for something it genuinely learned.
There is no internal flag separating recall from construction, which is why the tone never changes.
The failure concentrates in predictable places, and knowing them is most of the defence.
Facts that appeared rarely in training are stored weakly and reconstructed rather than retrieved.
Precise numbers, dates, identifiers and quotations are the worst case, since near enough is still wrong.
Anything after the knowledge cutoff cannot be known, so the model fills the gap by extrapolating.
Long chains of reasoning compound the risk, because one invented step is carried forward as established.
Retrieval-augmented generation is the main mitigation: put the source text in the context and ask for grounding.
That converts a memory problem into a reading problem, which models are far more reliable at.
Requiring citations that point at supplied passages makes the failure visible rather than invisible.
Measuring it needs faithfulness and groundedness checks rather than general quality scores.
It cannot currently be eliminated, so any system where a wrong answer is expensive needs a human or a verifier in the loop.