Reading · 7 min · Lesson 3 of 8
Hallucinations and how to catch them
Why models invent facts, and simple habits that protect you.
A hallucination is an answer that sounds right but is false: an invented statistic, a court ruling that never existed, a product feature your company does not offer. It happens because the model is optimised to produce plausible text, and when it lacks the facts it fills the gap with something that fits the pattern.
Hallucinations cluster in predictable places: exact numbers, dates, names, quotations, references and anything recent or specific to your organisation. General explanations and rewriting text you supplied are far safer, because the model works from material in front of it.
Three habits catch most problems. Give the model the source material and ask it to answer only from that. Ask it to say when it is unsure instead of guessing. And check every number, name and citation against the original before the text leaves your hands.
Key takeaways
- Hallucinations are fluent answers that are factually wrong.
- Numbers, names, dates and citations are the riskiest parts of any answer.
- Grounding the model in your sources and checking key facts prevents most errors.