Why AI "hallucinates" (and how to avoid it)
AI assistants sometimes invent facts with confidence. Understanding why, and adopting the right reflexes so you don't take them at their word.
Updated on
You ask an AI assistant for a reference. It cites a book, an author, a precise page. It all looks credible. Except none of it exists. Welcome to the world of hallucinations — one of the 10 key words of generative AI to know.
What is an AI hallucination?
A hallucination is a false answer stated with the same confidence as a correct one. The model does not lie in the human sense: it is not aware of being wrong. It produces what most resembles a good answer.
Why does an AI hallucinate?
A language model does not consult a database of truths. It predicts, word after word, the most plausible continuation of a text, from learned regularities. This is the very way a model works, which we break down in what an LLM is, explained simply. Most of the time, plausible and true coincide. Sometimes not: the model “fills the gaps” with something believable but inaccurate.
Hallucinations are therefore more frequent when:
- the question concerns a precise, rare or recent fact;
- you ask for a source, a figure, an exact quotation;
- the topic is niche and poorly represented in the training data.
The common forms of hallucination
They do not all look alike. Spotting the form helps you know what to check:
- The invented source: a credible book, article or URL… that does not exist. The easiest to unmask: open the link.
- The wrong figure or date: a precise statistic, a year, an amount — stated with confidence, false in the detail. The most dangerous, because precision inspires trust.
- The fabricated quotation: a sentence attributed to someone who never said it, or distorted.
- The recent fact: on an event after its training, a model with no web access fills the void instead of admitting it does not know.
A typical case, dissected
You ask: “What case law governs remote work for managers in France?” The model answers with a precise ruling — a court, a year, a number. It all looks plausible: it is the format of a real legal reference. But the ruling does not exist; the model assembled believable fragments to “produce an answer.”
What gave it away is not a glaring mistake, it is the excess of precision with no source. The reflex that would have avoided the trap: ask for the link to the ruling, then open it. An invented reference does not survive that check.
The reflexes that protect you
You cannot eliminate hallucinations, but you can greatly reduce the risk of being caught out.
Ask for sources, then check them. Do not settle for their apparent existence: open the link. An invented source does not survive a click.
Give the context instead of asking for it. If you provide the document, the model relies on it rather than on its memory. “Summarise this text” is safer than “what does the law say about…”.
Allow doubt. Add “if you’re not sure, say so.” A model allowed to answer “I don’t know” invents less.
Cross-check what matters. For a figure, a date, a legal or medical reference, a second human source remains indispensable.
To each risky situation, its reflex
In practice, a few kinds of request concentrate the risk. This table pairs each with the right reflex:
| What you ask for | Risk | The reflex |
|---|---|---|
| A source, a quotation, a reference | High | Open the link; no check, no trust |
| A precise figure or date | High | Cross-check with a reference source |
| A very recent fact | High | Check the model’s knowledge date, or provide the source |
| A summary of a text you provide | Low | Stay on the pasted document, do not broaden |
| A rephrasing, an idea, an outline | Low | Safe use: no external fact at stake |
The rule that emerges: the risk rises the moment external factual information is at stake, and drops when you work on content you provided yourself.
The right stance
Treat the assistant like a brilliant intern who is sometimes too sure of themselves: excellent for a first pass, never for validating alone what carries real stakes. Used this way, it saves real time. Taken at its word, it exposes you to costly mistakes.
Sources
Frequently asked questions
What is an AI hallucination?
It is a false answer stated with the same confidence as a correct one. The model does not lie in the human sense: it simply produces what most resembles a good answer.
Can you completely stop an AI from hallucinating?
No. You cannot remove hallucinations, but you sharply reduce the risk of being fooled by asking for sources, providing context, allowing doubt and cross-checking what matters.
When is the risk of hallucination highest?
When the question concerns a precise, rare or recent fact, when you ask for a source, a figure or an exact quotation, and when the topic is niche and poorly represented in the training data.
How can you reduce the risk in practice?
Ask for sources and open the links, provide the document rather than relying on the model's memory, add an instruction like 'if you are not sure, say so', and cross-check any sensitive figure, date or reference with a second human source.
Do recent models hallucinate less?
Generally yes, recent versions get common facts wrong less often. But hallucination is a structural limit, not a bug that gets fixed: it reappears on niche, rare or recent topics. Better models never excuse you from checking what matters.
Does an AI connected to the internet hallucinate less?
On recent facts, often yes: searching before answering reduces some inventions. But the AI can misread a page or attribute to it words it does not contain. Search shifts the risk, it does not remove it: always open the cited source.