Why AI makes things up
An AI chat writes by predicting which words should come next, based on the huge amount of text it learned from. Most of the time that gives good answers. But unless the app searches the web for you, it is answering from memory. Even when it does search, it can misread a page or mix two pages together.
When it does not know something, it may fill the gap with something that sounds right. People call this a hallucination. A made-up fact looks just like a real one: same calm tone, same neat sentences. So you cannot judge an answer by how sure it sounds.
Where mistakes hide
Some parts of an answer are much riskier than others. Check these first:
| Check this | Why |
|---|---|
| Names of people, books, studies and companies | AI can invent a real-sounding title or mix up who did what. |
| Numbers, prices and statistics | Figures can be old, mixed up or made up. |
| Dates and deadlines | AI can mix up years, and what it knows stops at some point in the past. |
| Links and sources | A web address can look real and lead nowhere, or to a page that says something else. |
| Quotes | A quote can be made up or credited to the wrong person. |
| Laws, rules, health and money facts | These change and differ by place. A mistake here costs the most. |
Ask in a way that shows the doubt
Asking "are you sure?" is not enough. The AI may just say yes, or change a right answer to a wrong one to please you. A better move is to ask, up front, for honesty about what it does not know.
Before
what are the rules for renting out my spare room?
After
What rules should I look into before renting out a spare room in Austin, Texas? Explain each one in plain words. For each rule, tell me which office or official website would have the real answer. If you are not sure about something, say so instead of guessing. Do not make up links.
No prompt can stop every mistake. This one asks the AI to show its doubts and points you to where the real answer lives.
Already have an answer? Paste it into a new chat and ask for a fact list to check. This does not prove anything by itself. It gives you a short list of what to check, starting with the shakiest parts.
You are a careful fact-checker. Here is an answer I got from an AI: [paste the answer] List every fact in it that matters: names, numbers, dates, quotes and links. For each one, say how sure you are (sure, fairly sure or not sure) and what kind of source I could use to check it, like an official website or the original document. Do not add new facts. If you think something is wrong or made up, say so plainly.
How to check, step by step
- Open every link. Make sure the page loads, comes from a real site you trust, and actually says what the AI claims. A link that works is not proof by itself.
- Search for names and titles. Type the exact book title, study name or person into a search engine. If nothing real comes up, treat it as made up.
- Check numbers at the source. Find the number on an official or original page, like a government site, the company's own price page or the report itself. Make sure it is current.
- Check dates against today. Ask the AI when its information ends. Anything that could have changed since then, like prices, rules or news, needs a fresh check.
- Use a second source. For anything important, find at least one source that did not come from the AI. If two good sources agree, you can trust it much more.
- Ask a pro for the big stuff. For health, legal, tax or money choices, use AI to understand and prepare, then confirm with a doctor, lawyer, accountant or other licensed pro.
Habits that keep you safe
- Give it the facts yourself. When you paste the real document and ask the AI to use only that, it has less room to invent. See how to give AI the right context.
- Ask it to say "I don't know". Add "If you are not sure, say so" to any prompt where facts matter.
- Match the checking to the stakes. A dinner idea needs no checking. A contract, a health question or a post your customers will read needs a careful look. If you run a business, like a law firm, check every fact before it reaches a client.
- Fix it in the chat. When you find a mistake, tell the AI what was wrong and ask it to redo the answer. More ways to push back are in follow-up prompts.
A simple rule: the more a wrong answer would cost you, the more checking it deserves.