AI hallucinations: a five-question habit for catching errors
Why AI assistants invent facts and citations, which outputs are most at risk, and a five-question review habit that catches most errors before they cause damage.
By the AiTrain editorial team · Reviewed under our editorial standards
An AI "hallucination" is output that sounds right and is wrong: a statistic nobody published, a court case that never happened, a product feature that does not exist. It is the single most important limitation for anyone using AI at work, because the error arrives in the same confident tone as everything else.
Key takeaways
- Hallucinations come from the way models generate text, not from a bug that will simply be patched away.
- Risk is highest for specific facts: numbers, names, dates, quotes, citations, laws and recent events.
- A short, consistent review habit catches the large majority of problems.
- Giving the model your source documents and asking it to quote from them reduces errors sharply.
Why it happens
A language model predicts likely text. When you ask for "three studies showing X", the most plausible-looking answer contains three study titles with authors and years — so that is what it writes, whether or not those studies exist. The model is not lying; it has no built-in sense of the difference between "I know this" and "this would fit here".
The risk map
| Output type | Risk | Why |
|---|---|---|
| Rewriting your own text | Low | The facts come from you |
| Summarising a document you pasted | Low–medium | May drop nuance or merge points |
| General explanations of well-known topics | Medium | Usually right, sometimes outdated |
| Statistics, prices, dates | High | Specific and easy to invent |
| Citations, links, quotes | Very high | Format is easy to fake |
| Legal, medical, financial specifics | Very high | Errors are costly and rules change |
The five-question habit
Before AI-assisted work leaves your hands, ask:
- Where did each fact come from? If a number, name or claim did not come from you or a source you can open, treat it as unverified.
- Can I open every citation? Click every link. Search every paper title. If it does not exist, remove it.
- Is anything time-sensitive? Prices, regulations, leadership roles and product features change. Check the current official source.
- Did it answer the question I actually asked? Fluent answers sometimes drift to an easier, adjacent question.
- Would I sign my name to it? If you could not defend a sentence to your manager or client, rewrite or cut it.
Prompts that reduce hallucinations
- "Use only the information in the document below. If the answer is not in the document, say 'not stated'."
- "Quote the exact sentence from the source that supports each point."
- "List anything in your answer you are not certain about."
- "Do not include statistics unless they appear in the material I provided."
These do not make the model perfect, but they change its default from "produce something plausible" to "stay inside the evidence".
A worked example
A marketing manager asked an assistant for "five statistics about remote-work productivity with sources". The answer listed five neat figures with publisher names. Checking took eight minutes: two figures matched real reports (with slightly different numbers), one came from a real publisher but a different year, and two could not be found at all. The fix was simple: she searched for two credible reports herself, pasted them in, and asked the assistant to pull the key figures from those reports only. The final result was accurate and still much faster than writing from scratch.
Building it into team practice
- Add a line to your AI use policy: "All facts, figures and citations in AI-assisted work must be verified against a primary source."
- Keep a shared list of near-misses. Real examples teach faster than rules.
- For high-stakes work, have a second person review sources, not just wording.
Teams that adopt this habit early tend to trust AI more, not less — because they know exactly where it needs watching.