Understanding hallucination.
WHAT IT IS
A model producing confident, fluent, false information: invented citations, fabricated statistics, non-existent products.
WHY IT HAPPENS
The system generates text that is statistically likely, not text that is true.
It has no internal distinction between recalling a fact and constructing a plausible sentence.
When it lacks information, it produces something shaped like an answer.
WHY IT SOUNDS CONVINCING
Fluency and accuracy are separate properties. The model optimises for the first.
A fabricated citation looks exactly like a real one because it was generated from patterns of real ones.
WHERE IT IS MOST DANGEROUS
Specific facts: names, dates, figures, quotations, references, legal provisions, medical dosages.
WHAT REDUCES IT
Providing the source material and asking about it Asking for uncertainty to be stated Using systems that search and cite Verifying anything specific
WHAT DOES NOT REDUCE IT
Asking the model whether it is sure. It will assert confidence either way.
THE RULE
Never publish a specific fact from AI output without checking it independently.
