AI hallucinations: why AI makes things up and how to protect yourself

AI makes things up because it produces the most plausible text rather than the most accurate one, and it sounds just as confident whether it is right or wrong. Protecting yourself comes down to a few simple habits: give it your sources, check anything precise, and keep sensitive topics for official sources.

What is an AI hallucination?

An AI hallucination happens when an assistant states false or invented information as if it were certain. The problem is not the mistake itself, since everyone gets things wrong, but the tone: nothing in the answer signals doubt.

Some typical examples:

• you ask for the biography of a little-known person, and the AI credits them with an invented degree or date of birth;
• you look for references for an essay, and it cites papers with believable titles that do not exist;
• you ask about a specific law, and it gives the wrong section number and content;
• you hand it a document, and it “finds” a sentence that is not there.

Why AI makes things up: three causes

It predicts what is plausible, not what is true

An assistant like ChatGPT runs on a large language model (LLM), which writes its answer piece by piece, each time picking the most likely continuation. It does not look anything up in an encyclopedia. When information is rare, ambiguous or missing from what it learned, it fills the gap anyway with whatever “sounds right”.

Training rewards guessing

In other words, a model that answers “I don’t know” often scores worse than one that takes a shot, like a multiple-choice test with no penalty for wrong answers.

It has gaps in its memory

Every model has a knowledge cut-off date: it knows nothing about what happened after its training ended, unless it can search the web. Niche topics, little-known people and documents too long to be read in full are also fertile ground for invention.

For example, you ask for the name of the new manager of your local sports club. That information has probably never appeared in the texts the model read. Rather than admit it does not know, it may come up with a believable name and an imaginary career, presented without the slightest caveat.

Whether AI “knows” it is wrong ties into a wider debate, which we cover in does AI really understand what it says?.

Where the risk is highest

Not all requests are equal. The more precise and checkable the expected answer, the more likely and costly a mistake becomes.

Type of request	Risk of invention	What to do
References, quotes, links	High	Open every link, search every title
Figures, statistics, prices	High	Trace the figure back to its original source
Laws, regulations, official procedures	High	Check the relevant government website
Little-known people, small businesses	High	Cross-check with an official site or directory
Recent news	Medium to high	Use an assistant with web search
Summary of a document you provide	Medium	Ask for exact quotes from the text
Rewording, drafting, ideas	Low	Check the substance, not just the style

Six habits to protect yourself

1. Provide the source yourself

An AI working from your document, contract or meeting notes makes fewer mistakes than one drawing on its memory. That is the idea behind grounding: basing the answer on information supplied at the time of the question. Paste the text or attach the file, then ask it to answer “using only this document”.

2. Turn on web search and open the links

For recent information, choose an assistant that offers AI web search and shows its sources. Perplexity, for instance, is built around this approach. But a link proves nothing until you have opened it: check that the page exists and actually says what the AI claims.

Assistant	Free	Paid from	Web search	Agent mode	Voice	Images
ChatGPT	Limited	$8/mo	Yes	Yes	Yes	Yes
Claude	Limited	$20/mo	Yes	Yes	—	—
Gemini	Limited	€4.99/mo	Yes	—	Yes	Yes
Le Chat (Mistral Vibe)	Limited	$5.99/mo	Yes	Yes	Yes	Yes
Perplexity	Limited	$10/mo	Yes	Yes	—	Yes
Microsoft Copilot	Limited	$9.99/mo	Yes	—	Yes	Yes

3. Allow “I don’t know”

Anthropic’s documentation on reducing hallucinations (accessed 22 October 2026) recommends explicitly giving the model permission to admit uncertainty. Simply add a line like this to your prompt: “If you are not sure or the information is missing, say so clearly instead of guessing.”

4. Ask for exact quotes

With a long document, first ask the AI to pull out, word for word, the passages that answer your question, then to write its answer from those extracts alone. If it cannot find a passage to back up a claim, the claim should go. The same documentation recommends this technique too.

5. Ask the question a different way

Ask the same question twice, or put it to two different assistants. If the answers disagree on a name, a date or a figure, that is a red flag: at least one of them is wrong.

6. Keep sensitive topics for official sources

For health, legal, tax or money matters, AI can help you understand the jargon or prepare your questions. The decision itself should rest on an official website (such as GOV.UK, the NHS, USA.gov or the IRS) or on a professional: a doctor, a lawyer, a financial adviser.

A step-by-step example

Say you want to know how much notice your lease requires before you move out. The risky approach is to ask “How much notice do I need to give my landlord?”. The AI will give you a general answer that may be right, may be out of date, and has nothing to do with your actual contract.

The much safer approach is to attach the lease and write an instruction along these lines:

Here is my lease. Answer using only this document. Quote word for word the clause about the notice period for moving out, then explain it in two simple sentences. If the document does not mention a notice period, say so clearly instead of guessing.

You now get an answer backed by a quote you can find in the contract in seconds. And if the clause seems unclear, you will know exactly which sentence to show your landlord, a tenants’ advice service or a lawyer.

This small effort combines three habits from the list: providing the source, asking for exact quotes, and allowing “I don’t know”.

Check yourself An AI cites three scientific papers with titles, authors and years. Can you use them as they are?

No. Highly specific references can be entirely made up. Search for each title on the journal’s website or in a database such as arXiv before citing them.

Warning signs to watch for

Some clues should trigger an immediate check:

• a very detailed answer on a very specialised or very local topic;
• suspiciously round or overly precise figures with no source;
• a source cited without a link, or a link that goes nowhere;
• a changed story when you ask “are you sure?”;
• a summary of your document containing information you do not recognise.

Going further

If you are new to the topic, our guide to what artificial intelligence is covers the basics, and our glossary of the 10 AI words you need to know demystifies the jargon. To write more precise requests that leave less room for invention, try the site’s prompt generator.