AI agents and technology
What is an AI hallucination? Definition and examples
Also: Confabulation
In short
A hallucination is a statement by an AI that sounds convincing but is false or unsupported. Language models produce likely sentences, not checked facts. When the right context is missing, they fill the gap with invention.
Why it matters: Hallucinations show where your AI lacks context.
Why does Hallucination matter for your company?
A language model writes, word by word, what is most likely to come next. Usually that is right; sometimes it produces a source, a clause or a number that does not exist. The danger is the tone: what is wrong sounds just as sure as what is right.
In companies some errors come not from invention but from outdated material: the AI correctly quotes a document that no longer applies. You reduce both the same way, with valid context and sources. Where errors are costly, a person checks as well.
Example: The invented customer commitment
An account manager asks an AI assistant to summarise the agreements with a key customer. The assistant names a discount commitment for 2027 that was never made. It sounds plausible, because similar commitments exist with other customers.
The team changes two things. The assistant may only state commitments with a source and otherwise writes “not supported”. And a person checks every summary that goes to customers before it is sent.
3 common mistakes to avoid
Waiting for a better model
Newer models hallucinate less often, but they still do. Without your context even the best ones guess.
Not asking for sources
An answer without a source cannot be checked. Ask for evidence, especially for numbers, commitments and legal questions.
Checking everything equally
A brainstorm needs no approval, a customer offer does. Check by consequence, not by gut feeling.
Step by step
- 1
Provide valid context
The more your AI reads from maintained sources, the less it has to guess.
- 2
Require sources
Factual statements come with a source or with the note “not supported”.
- 3
Check by consequence
A person checks anything that leaves the company or costs money.
- 4
Report errors
Record where the AI was wrong and close the gap in its context.
Bring one important decision.
In 30 minutes with Marc or Simon we show where your work drifts from the decision today.
Meet the Company BrainRead on in the PM Lab
Articles that go deeper.
Related terms
- RAG →RAG (retrieval-augmented generation) is a method where an AI model looks up relevant passages in your documents before it answers.
- Context window →The context window is the amount of text an AI model can read at once when it answers.
- Human in the Loop →Human in the loop means a person reviews and approves an AI process at defined points before it continues.
- AI governance →AI governance is the set of rules a company uses to decide who may use AI for what, with which data and under whose responsibility.