What is an AI hallucination?

You ask an AI a question.

It answers fast, clearly, in a confident tone.

Then you check… and the reference doesn’t exist. The quote was made up. The number came from nowhere.

It isn’t malice. It’s what people call a hallucination.

In plain words

An AI hallucination is a false or invented answer, presented as if it were true.

The AI may:

• invent a book, an article, or a study;
• fabricate a quotation;
• invent a number, a date, or a name;
• mix real facts with made-up details.

The trap: it often does this with the same confidence as a correct answer. No little sign that says “warning, I’m inventing.”

Why does it invent?

An AI like ChatGPT doesn’t “know” things the way a person who checked their sources does.

It predicts the next word (or rather the next small piece of text, a token) that seems most likely, based on what it learned.

Picture someone finishing a sentence out loud:

“The cat is sitting on the…”

The most likely word is often “mat” or “sofa.” Not necessarily the truth about your living room.

The AI does something similar, at a large scale:

• it chains words that “sound right”;
• it doesn’t automatically open a book, an official site, or your folder to verify.

Result: a smooth sentence can be… completely made up.

A good prompt helps a lot. But even a clear request doesn’t always prevent a hallucination: the tool remains a predictor, not a fact-checker.

Everyday examples

Nothing scientific here — just situations people run into often.

The fake reference

You ask: “Give me a serious article on this topic.”

The AI replies with a title, an author, a year… and the link goes nowhere. The article never existed. The names “fit” well together, so it produced them.

The invented quote

You ask for a quotation from a well-known person.

The AI gives you a nice one, with quotation marks and an attribution. That person never said it. The line sounds like their style — that’s enough for the machine.

The invented number

You ask: “What’s the average cost?” or “What’s the exact date?”

The AI gives a precise number, sometimes with decimals, as if it had looked at a table. In reality, it produced a plausible figure, not necessarily the right one.

In all three cases, the tone is confident. That’s why it fools people.

How do you spot the problem?

A few simple habits:

1. Ask for sources. “Where does that come from?” If it invents references, that’s a red flag.
2. Check elsewhere. Open the link. Search for the title. Compare with an official site, a document you know, or a trusted person.
3. Be wary of over-precise claims with no proof. An exact percentage or a day-level date, without a source, deserves a second look.
4. Use your common sense. If it sounds too good, too neat, or too surprising… pause.

You stay the judge. The AI suggests; you validate.

How do you limit inventions?

You can’t wipe them out completely. You can reduce them.

• Give the context yourself. Paste the text, the email, the procedure. The less it has to invent, the better.
• Ask it to stick to your documents. For example: “Answer only from the text below. If you can’t find it, say so.”
• Use a system that consults your files. That’s the idea behind RAG: before answering, the AI looks things up in a “library” (your notes, your policies, your manuals). It doesn’t make the AI infallible, but it anchors the answer in real information.

👉 What is RAG?

• Keep a human in the loop for anything that matters: health, money, legal questions, important decisions.

🍳 The simple version

Remember this:

A hallucination is when the AI invents — often in a very sure tone.

It predicts words that sound good. It doesn’t check on its own.

Your role:

• ask for sources;
• verify;
• give it your own documents when you can.

AI remains a useful tool. With a little care, you stay in control — and you avoid following a castle in the clouds. 🤖🔍