Hallucination — Fluent but False
A customer asked whether their wheels were covered by a warranty. The drafted reply said: "Yes! Spoke & Chain offers a lifetime warranty on all wheels." It was polite, well written and completely invented. The shop has a one-year warranty on labour only. This is a hallucination, and it is the most important failure to understand.
The rule
A model always sounds sure; only supplied facts make it right.
A hallucination is output that reads as confident and fluent but is not supported by facts. It is not a rare glitch. It follows directly from how the model works: it produces the most plausible continuation, and plausible is not the same as true. When the right fact is not available, the model does not stop; it writes what a correct answer would usually look like.
Hallucinations get more likely when:
- the question needs specific facts the model was never given (policies, prices, order details);
- the question is about rare or recent topics;
- the prompt pressures an answer ("Give the exact figure") with no permitted way to say "I don't know";
- the task needs exact values such as citations, numbers or quotes.
The confident tone tells you nothing. A hallucination and a correct answer sound equally sure.
The shortcut
Use the ground, limit, allow pattern:
- Ground — paste the facts into the prompt.
- Limit — tell the model to use only those facts.
- Allow — give it an approved way to say it does not know.
Here is our warranty policy:
<policy>
Labour on repairs is guaranteed for one year. Parts carry the maker's warranty.
</policy>
Answer using only the policy above. If it does not cover the question,
say "I'll check with the team and get back to you." Do not guess.Spot the gap
The prompt asked about warranties but contained no warranty facts.
Grounding reduces hallucination a great deal; it does not remove it. For anything that matters — money, safety, legal promises — check the output against the source.
Drills
Answer fast — trust the rule.
What is a hallucination?
Why do models hallucinate?
Which prompt is most likely to cause a hallucination?
Which part of "ground, limit, allow" gives the model a safe exit?
Remember
- A model always sounds sure; only supplied facts make it right.
- Shortcut: ground the facts, limit to them, allow "I don't know".
- Trap: trusting a confident tone — check anything that matters.