What Is an AI Hallucination? 5 Common Examples and How to Catch Them

You ask an AI assistant a question and get a clear, confident, well-written answer. Then you discover part of it is simply not true. This is called an AI hallucination, and it is the single most important thing to understand before you rely on AI at work.

Why AI makes things up

AI chatbots like ChatGPT, Gemini, Claude and Copilot are built on large language models. They write answers by predicting what a good reply is likely to look like, based on enormous amounts of text they learned from.

Most of the time, what looks right is also right. But when the model does not actually know something, it does not always say "I don't know". It can fill the gap with something that sounds plausible. The writing stays fluent and confident, which is what makes these mistakes easy to miss.

5 common kinds of hallucination at work

1. Sources that do not exist

You ask for references and get neat citations: a report title, an author, a link. Some of them may not exist, or may not say what the AI claims.

Catch it: open every link and find the exact line. If you cannot find it, treat the claim as unverified.

2. Numbers that are slightly wrong

Totals, percentages and growth rates in summaries can be off, especially when the AI calculates from a table you pasted. "Sales grew 18%" might really be 14%.

Catch it: recalculate every number in Excel or on a calculator before it goes into a report.

3. Old rules presented as current

Tax rates, deadlines, limits and procedures change. The AI's knowledge has a cut-off date and may mix up versions from different years, or from other countries.

Catch it: for anything legal, tax or regulatory, check the official source (for example, the GST portal, the income tax portal or the official circular) or ask a qualified professional.

4. Features that do not exist

Ask how to do something in software, and the AI may describe a menu, button or function that your version does not have, or that never existed.

Catch it: tell the AI your exact software and version, and confirm in the app or its official help. If the option is missing, say so and ask again.

5. Confident details about people or small companies

Ask about a specific person or a small business, and the AI may produce details that sound right but are invented, because there was little reliable information to learn from.

Catch it: do not use AI as a source of facts about specific people or companies. Use official records and the company's own website.

Warning signs

Slow down and check when an answer:

• Contains exact figures, dates or percentages
• Mentions laws, rules, rates or deadlines
• Gives sources, links or quotes
• Is about something recent
• Is about a specific person or a small organisation
• Is exactly what you hoped to hear

A 5-step checklist before you rely on an answer

1. Recalculate numbers yourself.
2. Open every source and find the exact line.
3. Check rules on official websites, not just in the chat.
4. Ask the AI where it might be wrong:

Review your answer above. Which parts are you least sure about, and what
should I verify from an official source before relying on it?

5. Ask an expert for anything high-stakes: money, legal, medical or tax decisions.

Make hallucinations less likely

You cannot remove the risk completely, but you can reduce it a lot:

• Give the AI the facts. Paste the actual document or data and ask questions about that, instead of asking from memory.
• Say "based only on this text". It keeps the AI from mixing in general knowledge.
• Allow uncertainty. Add "If you are not sure, say so instead of guessing."
• Ask for quotes and reasoning. "Quote the line you rely on" and "show the calculation" make errors easy to spot.

The bottom line

AI is an excellent assistant for drafting, explaining and organising. It is not a reliable source of facts on its own. Use it to work faster, and keep the checking habit for anything that matters.

Lesson 5 of our free course AI for Everyone covers fact-checking in more depth, with a practice task you can do in ten minutes.