Why AI Makes Things Up, and How to Check What It Tells You

You ask a chatbot for a book recommendation, and it invents a book that does not exist. Or it gives a detailed legal rule that is wrong. This is called a hallucination, and it is one of the most important things to understand about AI.

What is a hallucination?

A hallucination is when an AI model produces something that sounds believable but is false: a made-up fact, a fake quote, a non-existent web link, or an invented source.

Why it happens

Because a language model predicts likely text instead of looking up verified facts, it can produce a confident, fluent sentence even when it does not know the answer. It is built to give an answer, not to say "I'm not sure". Hallucinations are more likely when:

• The topic is obscure or very recent.
• You ask for exact numbers, quotes, citations, or links.
• The question includes a false assumption.
• The conversation is long and confusing.

A 5-step habit to check AI answers

1. Ask for sources, then open them

   Ask the AI to give links, and click each one. If a link is broken or the page does not say what it claims, do not trust it.

2. Cross-check key facts

   Look up the main claims on a reliable site such as a government, university, or respected news source.

3. Double-check numbers, dates, and names

   These are the details most often wrong.

4. Ask it to explain its reasoning

   Then look for steps that do not make sense. You can also ask "How confident are you, and what could be wrong?"

5. Test it with a second tool or a human expert

   For health, legal, money, or safety matters, ask a professional.

Ways to reduce errors

• Give the model the source text yourself and ask it to answer only from that text.
• Ask clear, specific questions.
• Use tools that connect AI to live search and show their sources, then still verify.
• Tell it that "I don't know" is an acceptable answer.

Where AI is safe and where to be careful

• Usually fine: brainstorming, rewriting, summarising text you provide, explaining concepts.
• Be careful: medical advice, legal or tax decisions, financial choices, exact citations, and anything where an error would hurt someone.

What most people miss

• Even experts have been caught out. Lawyers have faced penalties for submitting court filings with fake cases made up by AI.
• Models are improving, but not perfect. Newer versions hallucinate less, but none are error-free.
• Confidence is not a signal. The tone of an AI answer is the same whether it is right or wrong.

Sources

• NIST: AI Risk Management Framework
• Google for Developers: Machine learning crash course