Term · AI
AI hallucination
Sometimes AI gives an answer that sounds very certain but is wrong. That is called a hallucination.
In short
An AI hallucination is an answer from an AI tool that sounds convincing but is incorrect or made up. For example, a source that does not exist or a wrong amount. It happens because a language model produces what sounds likely, not what has been checked to be true. That is why you check important facts yourself.
What is an AI hallucination?
A hallucination is an answer from AI that is wrong, even though it sounds convincing. The AI makes up a fact, a name, a legal article or a source. You often cannot tell from the text that something is wrong.
The word is a figure of speech. The AI does not see things that are not there. It makes a mistake that sounds like a fact.
How does a hallucination happen?
A language model produces text by predicting what is likely to come next. It does not look up whether something is true. If the model does not know the answer, it often still produces an answer that fits the pattern.
Researchers at OpenAI wrote in 2025 that the way language models are trained and tested rewards guessing. A guess scores better than admitting you do not know. According to OpenAI, hallucinations occur less often in newer models, but they do not disappear.
The chance is greater with details that rarely occur: exact figures, dates, names, quotes and sources.
Example in an SME
A made-up example
An employee asks an AI tool how long you must keep personnel files. The answer cites a legal article and a period, and it sounds right. When she checks, that article number turns out not to exist. Because she checked it, it did not go out to her colleagues.
How do you check an AI answer?
- Check the facts. Look up amounts, dates, names and legal articles at the source itself.
- Ask for sources, and open them. A source can be made up too. Click on it and read along.
- Provide the right information. Put the text that should contain the answer in your prompt. Or use a system with RAG, which searches your own documents first.
- Let a person decide. Use AI text as a draft, especially for customers, contracts and money.
Knowing where AI goes wrong is part of AI literacy. In an AI training your team practises recognising and checking such mistakes, using its own work.
Frequently asked questions
How often does AI hallucinate?
That differs per model, per tool and per type of question. There is no fixed figure that applies to your work. Assume that it can happen, and check what matters.
Can you prevent hallucinations?
You cannot prevent them completely. You can make them less likely: give the AI the right information, ask for sources and have a person check the answer.
So does AI lie?
No. Lying is knowingly saying something untrue. A language model has no intention. It produces a text that sounds likely, and that text can be wrong.
Sources: OpenAI: Why language models hallucinate (5 September 2025), Microsoft Learn: Data, Privacy, and Security for Microsoft Copilot and Call from the ACM and AP on AI chatbots (in Dutch). Consulted on 7 October 2026.
Written by Bart van Engelenhoven Founder of Uniekaiconsultancy in Ede, the Netherlands. Bart analyses how SMEs work and builds software, integrations and AI applications himself. Last updated: 7 October 2026
Read more
Term
LLM (large language model)
An LLM (large language model) is the language model behind services like ChatGPT. It keeps predicting the next word. What that means for your business.
Read moreTerm
Prompt
A prompt is the instruction you give an AI tool. What do you put in it (task, context, form) and which data do you leave out? Explanation with an example.
Read moreTerm
RAG (retrieval-augmented generation)
RAG lets a language model search your own documents first and then answer, with a reference to the source. How it works and what to look out for.
Read moreNext step
Curious what could work for you?
- The first conversation is free and without obligation
- By phone or at your premises
- You speak directly with me
Bart van Engelenhoven You tell me what takes up your time, I tell you honestly whether I can help.