Why AI makes things up, and what you can do about it
AI can give a confident answer that is completely wrong. Here is why that happens, how to spot it, and when you should check the answer yourself.
You ask an AI chatbot for a book recommendation. It gives you a title, an author, and a neat summary. There is one problem: the book does not exist.
People often call this a hallucination. The simpler explanation is that the chatbot produced a sentence that looked like a good answer. It did not first check whether every fact was true.
- Start with what the model is actually doing
- Notice when the risk is higher
- Ask for evidence you can open
Start with what the model is actually doing
A language model builds an answer by predicting useful pieces of text from the words and context it received. This is why it can explain an idea clearly, rewrite an email, or continue a story.
Prediction is not the same as checking a database. If the model has weak information, it may fill the gap with something that sounds right.
Notice when the risk is higher
Wrong answers are more likely when you ask for obscure facts, recent events, exact quotes, links, or details about a person with little public information. Long and complicated questions can also hide a bad assumption.
The cost of an error matters. A made-up film title is annoying. A made-up dosage, contract clause, or bank detail can cause real harm.
| Option | Good fit | Trade-off |
|---|---|---|
| Brainstorming | Use the answer as a starting point | Ideas can still be generic |
| Facts and sources | Open and check the original source | Links and quotes may be invented |
| Health, law, or money | Use a qualified professional | A fluent answer can still be unsafe |
Ask for evidence you can open
Ask the chatbot to separate what it knows from what it is unsure about. Ask for sources, then open them. Check that the source says what the answer claims.
For your own documents, use a system that shows the exact page or section behind an answer. A citation is useful only when a person can follow it.
Give it permission to say I do not know
Many prompts accidentally reward confidence. Say that an uncertain answer should be marked as uncertain. Tell the model not to invent missing names, dates, or links.
This will not remove every mistake. It does make honest uncertainty more likely and easier to notice.
Use AI for the part it does well
AI is excellent for drafting, organising, comparing options, and helping you see a problem from another angle. Treat important factual claims as work that still needs checking.
A good habit is simple: use the chatbot to move faster, then slow down wherever a wrong answer would matter.
How we prepared this guide
This guide explains the topic in plain language using current primary documentation and practical examples. It does not claim hands-on benchmarking. Check the linked sources for current product details.
Facts reviewed
