A hallucination is the model stating something completely false with total confidence. A citation that does not exist, a policy you never wrote, a number that was never in any report, delivered in the same fluent, assured tone as everything true the model says. There is no stammer, no flag, no drop in polish to warn you.
The uncomfortable part: it is not a glitch. The model predicts plausible text. It does not check truth, because it has no mechanism for checking truth. Confident and wrong is not the model malfunctioning, it is the model doing exactly what it was built to do on a question where plausible and true came apart. That framing matters, because you cannot patch a behavior that is the system working as designed. You can only build around it.
For an operator, the rule is simple: never let unverified model output touch a customer, a contract, or a number that matters. The guards are known and buildable. Ground the model in your real documents so it answers from evidence instead of memory. Keep a human review step wherever the cost of a wrong answer is real. Wrap the workflow in checks that catch the failure before it leaves the building.
A concrete example. A support assistant is asked about a refund policy the company never published. An ungrounded model will happily compose one, politely and precisely, and a customer will act on it. The same assistant, grounded in the actual policy pages and instructed to say when the answer is not there, replies with what is true or admits the gap. The second build treats hallucination as a design constraint, which is the only honest way to treat it.
“What's our guard against hallucinations before this touches a customer?”
// this page is the full text. the packaged PDF, all fifteen terms, comes with the free library.