How to Fact-Check AI: Spotting Hallucinations Before They Cost You
Date Published
AI's most dangerous trait is not that it gets things wrong. It is that it gets things wrong in the same calm, confident voice it uses when it is right. There is no nervous pause, no "I think", no smaller font for the parts it invented. Learning to catch those confident mistakes is the single most valuable AI skill an office worker can have, and it takes about five minutes to learn.
This guide explains why AI makes things up, and gives you a simple routine to catch it before a made-up number ends up in your report.
Why AI makes things up
When an AI does not know something, it does not stop. It predicts the most plausible-sounding continuation, which is often correct and occasionally pure fiction dressed as fact. The industry calls this a hallucination, and it is a feature of how these tools work, not a bug that will fully disappear. It shows up most with specific facts: statistics, dates, names, quotes, legal citations, and study references. The fix is not to distrust everything, but to know exactly where to look.
The high-risk list: what to always check
Not everything needs verifying. General explanations and drafting are usually safe; the danger lives in specifics. Treat these as guilty until proven innocent: any statistic or percentage, any named source or study, any date or deadline, any price or legal rule, and any quote attributed to a person. If you are about to repeat one of these to your boss or a client, it earns 30 seconds of checking.
Move 1: make the AI show its work
The fastest first filter is to ask the AI to separate what it knows from what it is guessing:
For the answer you just gave, go through each factual claim and label it: SOLID (widely established), LIKELY (probably right, worth checking), or UNCERTAIN (you are not sure). For anything not SOLID, tell me exactly what I should search to verify it. Do not defend the claims, just rate them honestly.
This will not catch everything, because a confidently hallucinated fact can be labeled SOLID. But it reliably flags the softer claims, and it turns a wall of assertions into a checkable list. The same labeling trick powers the verification step in our document summarizing guide.
Move 2: demand real sources, then open them
Ask for sources, but never trust the citation itself until you click it, because invented sources are a classic hallucination:
Give me the specific sources for the key facts in your answer, with links. For each one, tell me what that source actually says and how it supports the claim.
Then open the links. If a link is dead, the title does not exist, or the page does not say what the AI claimed, the fact is unverified. Using an AI tool with live web search built in helps here, because it cites pages that actually exist, though you should still open them for anything important.
Move 3: cross-check the one number that matters
You rarely need to verify a whole answer. You need to verify the one fact you are about to act on. Pull it out and check it against a source you already trust: the official website, the original document, a quick independent search. As tech journalist coverage of AI fact-checking keeps repeating, a second independent source is the whole game. If two trustworthy sources agree, you are almost certainly safe. If they disagree, you have found exactly the thing that would have embarrassed you.
A 30-second habit that saves careers
Before any AI-sourced fact leaves your hands, ask one question: would it hurt if this were wrong? If the answer is no, ship it. If yes, spend the 30 seconds. That single filter, applied consistently, is what separates people who use AI well from people who get burned by it. It pairs directly with the data-safety habits in our guide to using ChatGPT safely at work, and the vocabulary in our plain-English AI glossary fills in the rest.
AI is a brilliant, fast, confident assistant that is sometimes wrong. Treat it exactly as you would treat a talented new hire: grateful for the speed, and quietly checking the important numbers. Learn that balance in six short lessons with our free ChatGPT from Zero course.
Frequently asked questions
Why does AI make things up?
When an AI does not know something, it predicts the most plausible-sounding answer rather than stopping. That prediction is often right and sometimes fiction stated as fact. This is called a hallucination and it is inherent to how the tools work, so verification is a permanent skill, not a temporary workaround.
What kinds of AI answers should I always check?
Specifics: any statistic or percentage, named study or source, date or deadline, price or legal rule, and any quote attributed to a person. General explanations and drafting are usually safe; the danger is in precise facts you plan to repeat or act on.
How do I know if an AI is hallucinating?
Ask it to label each claim by confidence and to provide sources, then open the sources. Warning signs include oddly specific statistics with no source, citations that do not exist when you search them, and answers that stay confident even when you push back.
Are AI tools with web search more accurate?
They reduce made-up citations because they reference pages that actually exist, and they are better for current facts. But they can still misread or overstate what a source says, so open the important links yourself rather than trusting the summary.
Do I need to fact-check everything AI writes?
No. Verify what you will act on or repeat, especially numbers, names, and dates. A simple filter works: would it hurt if this were wrong? If yes, spend thirty seconds checking against a trusted source; if no, move on.
Sources
Yes, if you follow a few rules. What happens to what you type, five things never to paste, the 10-second anonymizing trick, and the privacy settings worth two minutes.
Paste, prompt, verify: a plain-English method for summarizing long reports with ChatGPT: the exact layered prompt, the chunk method for big files, and how to catch mistakes.