Ask an AI something it doesn't know and it may not say so. It may invent an answer instead. And it will sound just as sure.
Picture a dinner guest who can’t bear to say “I don’t know”. Ask them who invented the toaster and you’ll get a name, a date and a little story, all told with total confidence. Some of it might even be true.
AI can be like that guest. A hallucination is when an AI makes something up and presents it as fact. It might invent a statistic, a quote, a court case or a book that was never written.
Why it happens
An LLM writes by predicting what words usually come next. It’s very good at sounding right. But it has no built-in way to check whether something is true, so when it doesn’t know, it can fill the gap with something that simply fits.
How to spot one
- a source, link or quote you can't find anywhere else
- very precise numbers with no explanation of where they came from
- details that change when you ask the same question again
What it isn’t
It isn’t lying. The AI isn’t trying to trick you. It doesn’t know the difference between what it remembers and what it has made up.
It also isn’t rare. Newer tools hallucinate less, especially when they can search the web or read your own documents. None of them are perfect yet.
Where the comparison breaks down
A dinner guest usually knows when they’re bluffing. An AI doesn’t, which is why it sounds just as sure either way.
How the words fit together
You write a prompt. The LLM answers one token at a time. If it fills a gap with something made up, that’s a hallucination.
Before you use an AI answer for anything that matters, check it. Be extra careful with names, numbers, quotes and links, because those are the easiest things for it to make up.

When you'll hear it
“It hallucinated a source.” It gave a reference to an article or book that doesn't exist.
“Always fact-check the output.” Don't trust AI answers blindly. Check the important bits yourself.
“Ground it in our own documents.” Give the AI the real information to work from, so it has less reason to guess.
“How will we check the AI's answers before they reach customers?”


