Explainer
AI hallucinations: why does AI make things up?
A hallucination is when AI confidently writes something that is not true – a made-up figure, a non-existent legal clause or the wrong date. It is not a bug in the program but a trait of language models. The good news: it can be managed well.
In short
- The model writes what sounds likely – not necessarily what is true.
- What helps most is giving AI your verified documents and asking it to cite the source.
- For important steps, a person always decides.
On this page
Why hallucinations happen
A language model does not contain a database of facts that it searches. It builds an answer so that it sounds natural and fits the question. When it lacks the information it needs, it often “fills in” something that looks plausible.
This typically happens with precise details: numbers, dates, names, references to laws, or internal company rules the model has never seen.
What a hallucination looks like in practice
- An online shop’s chatbot promises a customer free delivery that you do not offer.
- An AI contract summary states a notice period that is not in the contract.
- A reply to a citizen quotes a local regulation that does not exist.
All three cases have one thing in common: the AI answered from its “general memory” instead of from your documents.
How to limit hallucinations
- Give AI your documents. The agent should answer only from the price list, policies or documents you approve. This technique is called RAG.
- Ask for the source. With every answer, the agent shows which document it drew on. Checking then takes seconds.
- Allow “I don’t know”. The agent has a clear instruction: if the answer is not in the documents, pass the question to a person.
- Check automatically whatever you can. Numbers, dates and company registration numbers can be verified against the system automatically.
- Let a person decide. Where money, terms or sensitive matters are involved, the agent only prepares a draft (human-in-the-loop).
Does this mean AI cannot be trusted?
No. It means AI should be brought in the same way as a new colleague: with clear documents, rules and checks. A well-built agent makes fewer mistakes because it does not work “off the top of its head” but from your data – and when it is not sure, it says so.
Frequently asked questions
Can hallucinations be prevented completely?
Not completely, but they can be greatly reduced: by working from your own documents, citing the source with every answer, checking details automatically and having a person approve important steps.
Do all models hallucinate equally?
No. Newer models hallucinate less and are better at admitting they do not know something. But how the AI is deployed matters more than which model you choose.
How can I tell that AI is hallucinating?
The most reliable way is to ask it for the source. If it cannot show where the information came from, treat it as unverified.
Want to know where AI would save you time?
We go through a normal week with you and find the first task worth handing to an agent. No obligation.
Related articles
RAG: how to teach AI your company documents
RAG is a way for AI to answer from your documents instead of its general memory. How it works, what it is useful for in a business and what to watch out for.
→Explainer · 3 minHuman-in-the-loop: a person decides
Human-in-the-loop means AI prepares the work and a person approves it. How to set an agent’s limits so it saves time without making costly mistakes.
→Explainer · 3 minLarge language model (LLM)
What a large language model (LLM) is, how it works and why it powers both ChatGPT and AI agents. A maths-free explanation for business owners.
→