This is the flaw that has cost professionals the most in recent years: a model inventing a reference, a date, a piece of legislation, a figure. The problem is not so much the error as the tone. The wrong answer arrives with exactly the same confidence as the right one.
The word "hallucination" is poorly chosen, in fact. It suggests an accident, an exceptional slip. It is really the system working normally, pushed into a zone where it has nothing solid to produce.
Definition
A hallucination is false but plausible information produced by an AI model, presented with the same degree of certainty as accurate information.
The cause is mechanical. A Large language model (LLM) is trained to produce the most likely continuation of a text, not to tell the truth. When the right answer is not in what it learned, it has no mechanism to stop: it keeps producing whatever most resembles a good answer.
A model has no internal confidence dial it could show you. It does not know that it does not know. This is why "are you sure?" achieves little: the answer to that question is produced by the same mechanism as the error.
Where the risk is highest
| High risk | Low risk |
|---|---|
| A precise verifiable fact: date, amount, reference | Rewriting a text you supply |
| A niche subject, thinly represented in the data | A very common, stable subject |
| A recent event, after training ended | A task about style or structure |
| A question asked with no context supplied | A question about an attached document |
The right-hand column outlines the main defence: the more material you supply, the less the model needs to invent. That is exactly the reasoning that produced RAG.
Three statements produced by a model. One is false. Which one?
How to protect yourself
Supply the source rather than asking for it. Pasting the document and asking for an answer grounded in it massively reduces the risk, compared with a question asked into the void.
Explicitly allow ignorance. An instruction such as "if the information is not in the supplied document, say you do not have it" changes a great deal. Without that permission, the model assumes it must answer.
Ask for quoted passages. Requiring every claim to point at an extract of the document makes checking fast, and makes inventions immediately visible.
Systematically verify precise details. Proper nouns, figures, dates, legal references, addresses. These are exactly the elements a model most readily produces in a plausible but wrong form.
Cross-check decisions that matter. Two different models asked separately rarely go wrong in the same way. On a high-stakes subject, that is a cheap control.
The most dangerous case is not the huge error, which is obvious. It is the discreet error in the middle of an otherwise excellent text: a slightly wrong figure, a date off by a year. The overall quality of the text lulls your vigilance.
Frequently asked questions
Do recent models still hallucinate?
Markedly less, and they admit ignorance more often than two years ago. But the phenomenon has not gone away and will not disappear entirely: it follows from the very principle of text prediction. A newer model moves the problem towards more specialised questions, it does not remove it.
Does internet access solve the problem?
It improves matters greatly on recent facts, but introduces another: the quality of the source found. A model citing an unreliable page produces a wrong answer with a reference attached, which is more convincing and therefore riskier.
Can you spot a hallucination by reading?
Rarely, and that is the whole problem. The style does not change. The only useful clue is statistical: the more precise and verifiable a claim, the more it deserves checking. A generality is unlikely to be invented, an exact reference very much is.
How do you build professional use despite this risk?
By reserving AI for tasks where you can judge the result, and keeping human review anywhere an error would leave your business. Our Claude Cowork course covers these guardrails alongside the gains, because the two are decided together.