What is an AI hallucination?

Why language models hallucinate

A large language model does not look facts up; it calculates which words are likely to follow. If the necessary knowledge is missing from the training data or the context, the model still produces a fluent answer – it fills the gap with something that fits linguistically. Typical causes are:

• Missing knowledge: internal company information or events after the training cut-off are unknown to the model.
• Unclear questions: vague or ambiguous prompts invite assumptions.
• Poor sources: outdated or contradictory documents in the context lead to wrong conclusions.
• Long contexts: with very large inputs, models can overlook or mix up details.
• Calculations: language models are not calculators; figures from tables are a frequent source of errors.

Why hallucinations are a business risk

A wrong figure in an offer, an invented contract clause or a non-existent source in a report can have real consequences – financial, legal or for your reputation. The tricky part: hallucinations are hard to spot because they are formulated just as confidently as correct answers. Teams therefore need clear rules on which results must be checked before they are used.

How to reduce hallucinations

1. Provide sources: connect the model to verified company knowledge via RAG and have it cite the passages used.
2. Ask precise questions: clear tasks, context and the explicit permission to say "I don't know" help – see prompt engineering.
3. Calculate instead of guessing: have figures computed with suitable tools rather than estimated from text.
4. Maintain knowledge: remove outdated documents and clarify responsibilities.
5. Keep humans in the loop: results with external impact – offers, contracts, customer communication – need review.
6. Use structured knowledge: relationships between customers, projects and products are more reliable in a knowledge graph.

What does not work

Hallucinations cannot be completely eliminated with today's technology. Neither a bigger model nor a perfect prompt guarantees correct answers. The goal is therefore not zero errors but verifiability: every important statement should be traceable to a source.

Hallucinations and AIPE

AIPE reduces the risk with source-based answers: documents are semantically indexed, retrieval only searches documents the user may read, and answers reference the source documents used, so statements can be checked. Spreadsheet questions are calculated with a dedicated analysis tool. AIPE cannot rule out hallucinations either – reviewing important results remains necessary. How to build a reliable knowledge base is explained in the guide AI knowledge base; the functions are described on the knowledge page.