AI hallucinations: why they happen and how to spot them

You ask an AI for a book on a specific subject. It gives you a title, an author and a year of publication, and it all looks right. Yet the book doesn’t exist. This is what we call AI hallucinations. They come from the way these tools work and are nothing unusual. Once you understand where they come from, you can learn to recognise them and limit them from the very first question.

What is an AI hallucination?

We talk about a hallucination when an AI produces false information and presents it as true. It might be an invented fact, a source that doesn’t exist, a figure out of nowhere or a quote nobody ever said. The word is vivid, and it describes the thing well. The tool can produce made-up information and present it confidently.

The AI has no intention of deceiving you. It simply can’t tell, from the inside, what it knows from what it’s piecing together.

Why AI makes things up

A conversational AI predicts, word after word, the most plausible continuation of a text. It learned from huge volumes of writing, which makes it very good at producing answers that look correct. A plausible answer can be right or wrong: its style isn’t enough to judge how reliable it is.

Hallucinations mainly appear in four situations.

• The information is missing, because the subject is very specialist, the event recent or the person little known.
• It’s asked for precision, such as an exact figure, a date or a full reference.
• The question steers the answer. If you ask “What’s X’s best book on Y?”, it may invent a book rather than say there isn’t one.
• The conversation is long, and it mixes up things mentioned earlier.

The most common hallucinations

Some mistakes come up often enough to be worth knowing.

• Invented sources, such as books, studies or links that don’t exist.
• Figures that are too precise, statistics or amounts with no clear origin.
• Quotes attributed to people who never said them.
• Dates, times and prices that are out of date or approximate.
• False biographical details about real people.
• Imaginary features in an app, with a button that doesn’t exist.

Warning signs to look out for

A hallucination doesn’t announce itself, but a few clues should make you suspicious.

• A very precise figure, to one decimal place, with no source.
• A reference you can’t find anywhere when you look it up yourself.
• A different answer when you ask the same question another way.
• An answer that fits perfectly on a subject where information is scarce.
• An AI that admits its doubt as soon as you ask “are you sure?”.

When one of these signs appears, use our 5-step method to check an answer.

Limiting hallucinations from the first question

You can also reduce the risk upfront.

• Give it the text. Asking it to summarise a document you provide is far more reliable than asking it from memory.
• Invite the tool to flag what it’s unsure about, for example with “If you don’t know, say so”. This instruction doesn’t replace checking important information.
• Ask for sources you can check, then open them.
• For recent topics, use web search when your assistant offers it.
• Avoid questions that assume the answer. Ask “Is there a book by X on Y?” rather than “What’s X’s book on Y?”.

Our guide on how to write a good prompt gives more tips for getting reliable answers.

Your profile changes how you spot them

We don’t all spot hallucinations in the same way. The Investigator can rely on their taste for finding sources and cross-checking information. The Experimenter spots them by comparing several tools. More trusting profiles gain from a few simple habits for what matters.

Our AI profile test measures your relationship with information, from verifier to trusting. In about 3 minutes, with no account, you’ll discover your profile and a first tip for using AI without being fooled.