Why AI Makes Things Up (and 6 Ways to Catch It)
AI 'hallucinations' are confident answers that are simply wrong. Here's why they happen, which questions trigger them and a simple routine to catch them.
Ask an AI chatbot for five books on a niche topic and you might get five great recommendations. Or three real books and two that sound perfect but don't exist, complete with plausible authors and publication years.
That's a hallucination: an answer delivered with total confidence that isn't true. It's the single most important thing to understand about AI tools.
Why it happens
AI chatbots don't look facts up in a database. They generate text that's likely to come next, based on patterns learned from huge amounts of writing. (We explain this in how AI chatbots actually work.)
Most of the time, the most likely text is also the true text. But when the model doesn't really know something, it doesn't go quiet. It produces something that looks like the right kind of answer: a realistic-sounding statistic, a citation in the right format, a confident date.
It isn't lying, because it has no intention to deceive. It's more like a very fluent person who'd rather guess than say "I don't know".
Questions that trigger hallucinations most
- Specific numbers: statistics, prices, dates, percentages
- Citations: paper titles, book references, legal cases, URLs
- Niche topics with little written about them
- Recent events after the model's training date (unless it searches the web)
- Details about people who aren't very famous
- Long chains of reasoning, where one wrong step carries forward
How to catch them: a six-step routine
1. Ask for sources, then check them
"Give me sources for that" is a good start, but click every link. A real, relevant page is a good sign. A broken link or a page that says something different is a red flag.
2. Use web-connected tools for facts
Tools that search the web and cite sources, like Perplexity or the search modes in ChatGPT and Gemini, are far less likely to invent facts about current events.
3. Give it the facts
Instead of asking the AI to remember something, paste in the document and ask about that. "Based only on the text below…" dramatically cuts down on made-up details.
4. Ask it to flag uncertainty
"Which parts of your answer are you least confident about?"
It won't catch everything, but it often points straight at the shaky bits.
5. Ask the same question differently
If you get two different answers to two phrasings of the same question, at least one is wrong. Dig in.
6. Verify anything that matters
Money, health, law, safety and anything you'll publish. Check with an official source, a professional or the original document. Every time.
Are hallucinations getting better?
Yes. Newer models make things up less often, and web search has helped a lot with facts. But they haven't disappeared, and the remaining ones are harder to spot precisely because the rest of the answer is so good.
The mindset that works
Use AI for thinking, drafting and explaining, where small errors are easy to catch and fix. Be skeptical with facts, figures and quotes. And never paste an AI-generated statistic into something important without checking where it came from.
