Home/Knowledge hub/What are AI hallucinations?
GlossaryWhat are AI hallucinations?
A language model always sounds convincing, also when it is wrong. That is a hallucination: an answer that looks right but is not based on facts.
In short
An AI hallucination is an answer from a language model that is plausible but incorrect, outdated or made up, because the model fills gaps instead of relying on facts. Hallucinations can lead to wrong decisions, compliance risks and a loss of trust. The most effective remedy is to give the model verified company knowledge as context, with sources, so that every answer can be checked.
On this page
Why do language models hallucinate?
A language model predicts the most likely next words based on what it learned. It does not check facts. When information is missing, outdated or ambiguous, it still produces a fluent answer. Typical causes:
- The answer is not in its training data, for example internal or recent information.
- The question is vague or lacks context.
- The documents it receives are outdated or contradict each other.
- Nobody asks for sources, so errors go unnoticed.
Risks for your organisation
Wrong decisions
Advice based on a contract clause that does not exist.
Compliance risk
Incorrect statements about rules, deadlines or obligations.
Reputation
Wrong answers sent to clients or partners.
Lost trust
Once people catch AI making things up, they stop using it.
How to prevent AI hallucinations
- Step 1
Ground answers in your data
Use RAG: retrieve relevant company documents first, then let the model answer from them.
- Step 2
Give relevant, verified context
Context Engineering selects a small set of pre-verified facts instead of whole documents.
- Step 3
Show sources
Every answer refers to the documents it is based on, so it can be validated easily.
- Step 4
Keep documents in order
Clear final versions, metadata and retention rules, so outdated information is not used.
- Step 5
Keep a human in the loop
Review answers with impact before they are used or sent.
Hallucinations, insight engines and Mindbreeze
According to Mindbreeze, a company can minimise hallucination by combining company knowledge extracted by an insight engine with the broader language understanding of a large language model. The insight engine provides the factual basis for the answer, so any answer can be validated easily.
Frequently asked questions about AI hallucinations
What is an AI hallucination?
An AI hallucination is an answer from a language model that is plausible but incorrect, outdated or made up, because the model fills gaps instead of relying on facts.
Can hallucinations be prevented completely?
Not completely, but the risk can be reduced strongly: ground answers in verified company data, show sources and keep a human in the loop for decisions with impact.
Does RAG stop hallucinations?
RAG reduces them, because the model answers from retrieved documents. The quality of the result still depends on the documents and on the relevance of what is retrieved.
How do I recognise a hallucination?
Check the sources. An answer without a source, or with a source that does not say what the answer claims, deserves a second look.
What is the role of document management?
Outdated, duplicate or unstructured documents are a major cause of wrong answers. Clear final versions, metadata and retention rules give AI a reliable foundation.
Our consultants are happy to think along with you.
Plan a conversationCan you trust the answers of your AI?
In Discovery we look at your documents and sources first, and tell you what AI needs to give reliable answers.
Continue learning
Related terms, articles and the solutions behind this topic.