Glossary
A hallucination is output that is fluent and confident but false: an invented citation, a policy that does not exist, a figure with no source. It is a normal consequence of how generative models work, not a malfunction.
There is no internal flag separating them, which is why the tone of a wrong answer is indistinguishable from a right one.
These are also, unhelpfully, the questions businesses most want answered.
Ground answers in retrieved source material, keep the corpus curated, show the source alongside the answer so a reader can check it, and define a confidence boundary below which the system escalates rather than guesses.
The realistic goal is to make errors rare, visible and recoverable, and to keep humans accountable for decisions that matter.
No. It can be made rare, visible and recoverable through grounding, source citation and escalation at low confidence. A supplier promising elimination is describing something they have not tested properly.
On questions about your own business that the model was never trained on, and on anything requiring an exact figure, date or reference. Those are the cases that most need retrieval and a visible source.
A hallucination is output that is fluent and confident but false: an invented citation, a policy that does not exist, a figure with no source. It is a normal consequence of how generative models work, not a malfunction.
