AI Hallucinations: Why They Happen and How to Check AI Output

An AI "hallucination" is output that sounds plausible but is false: an invented statistic, a quote nobody said, a court case that does not exist, a reference to a paper with a convincing title and no author who ever wrote it. The tone gives no warning. The answer reads just as confidently as a correct one.

You do not need to avoid AI tools because of this. You need to know when errors are likely and how to check efficiently.

Why hallucinations happen

A language model generates text by predicting what words are likely to come next, based on patterns learned from enormous amounts of training data. It is extremely good at producing text that looks like a correct answer. It does not, by default, look facts up or know which of its statements are true. When the training data is thin, conflicting or out of date, the model still produces fluent text, and that is where errors creep in.

When errors are most likely

• Specific numbers: statistics, prices, dates, percentages and measurements.
• Citations and quotes: titles of papers, URLs, page numbers and attributed quotations.
• Niche or local topics with little public information.
• Recent events after the model's training data ends, unless the tool searches the web.
• People's biographies, especially less famous people.
• Long chains of reasoning where one early mistake carries through.
• Leading questions. Ask "Why is X true?" and a model may justify X even when it is false.

A practical checking routine

1. Highlight every checkable claim

Before editing for style, mark each fact: names, numbers, dates, quotes, sources and causal claims. This turns a vague worry into a short list.

2. Verify against primary sources

Check numbers against the original publisher: the statistics office, the regulator, the company's own report, the research paper itself. A secondary article that repeats the figure is not confirmation.

3. Open every link and reference

Confirm the page exists, says what the text claims it says, and is the original source. Fabricated references often have realistic titles attached to real journals.

4. Ask the model to show its working

"Which parts of your answer are you least certain about?" and "What would change this conclusion?" often surface weak points. Treat the response as a lead to investigate, not as proof.

5. Cut what you cannot verify

If a claim cannot be confirmed quickly, remove it or rewrite it as clearly uncertain. Publishing nothing is better than publishing something false under your name.

Reduce errors before they happen

• Provide the source material and instruct the model to answer only from it. Retrieval-based systems do this automatically; see our guide to RAG.
• Allow "I don't know". Tell the model explicitly that saying it is unsure is acceptable.
• Ask neutral questions rather than ones that assume the answer.
• Use AI for structure and drafting, and bring the facts yourself. Our prompt-writing framework includes constraints that help.

Match the checking effort to the stakes

Be open about AI use

If AI helped produce something you publish, the responsibility for its accuracy is still yours. Many organisations now add a short note explaining how AI was used and how content was checked. Readers tend to accept AI assistance; what damages trust is discovering an error that nobody checked.