Key takeaways
- A hallucination is a plausible-sounding but false or unsupported AI statement.
- It occurs because language models generate likely wording, not verified knowledge.
- Your own sources, clear prompts and source citations reduce the risk.
- For important content, human review remains mandatory.
Contents
What is an AI hallucination?
A hallucination occurs when an AI model produces a statement that sounds plausible but is false or cannot be supported. Examples include invented figures, non-existent legal provisions, misattributed quotes or citations of sources that don’t exist.
The tricky part: hallucinations are linguistically indistinguishable from correct answers. The model phrases them with the same confidence. The OWASP list of the most important risks for LLM applications therefore explicitly includes misinformation as a risk of its own.
Why language models hallucinate
A large language model generates text by calculating likely continuations step by step. It doesn’t look things up in a verified database. A research survey on hallucinations names, among other causes:
- Gaps and errors in the training data: What the model never learned, or learned incorrectly, it cannot reproduce correctly.
- Outdated knowledge: The model only knows the state of its training data.
- How text is generated: The model is trained to produce fluent answers – even when it doesn’t actually “know”.
- Unclear requests: When context is missing, the model fills the gaps with guesses.
How language models work in principle is explained in What is an LLM?.
Where the risk is particularly high
- exact figures, dates, prices and statistics
- source citations, quotes and references
- legal provisions, rulings and standards
- rare specialist topics and niche knowledge
- current events after the model’s training cutoff
- statements about specific people or companies
Seven ways to reduce hallucinations
- Provide your own sources: Let the AI work with your documents instead of answering from memory. This approach is called retrieval-augmented generation; see RAG: using AI with company knowledge.
- Ask for citations: Request evidence with a location – and check that the source exists and actually supports the statement.
- Ask precisely: The clearer the task and context, the less the model has to guess. More in What is a prompt?.
- Allow uncertainty: Write in the prompt: “If you’re not sure about something, say so.”
- Split up tasks: Have complex questions handled in steps and check the intermediate results.
- Choose the right tool: For current information, an AI with web search is better suited than a model that only answers from its training.
- Define human review: Specify which content must be professionally reviewed before use.
A simple review rule for your team
Set out in your AI policy when content gets reviewed. A practical rule of thumb: the greater the consequences of an error, the more thorough the review.
| Use | Review |
|---|---|
| Ideas, outlines, internal drafts | Plausibility check by the user |
| Texts for customers and partners | Check facts, figures and commitments |
| Contracts, legal, finance, health | Professional review by the responsible person |
A template for such rules is in AI policy template.
Hallucinations and AI literacy
Since February 2025, Article 4 of the EU AI Act has required measures so that employees can use AI competently. The European Commission explicitly mentions awareness of risks such as hallucinations for everyone who uses chatbots. A short training session with examples from your own day-to-day work is a good start.
Frequently asked questions
Can hallucinations be prevented completely?
No. According to current research, they can be reduced significantly but not ruled out entirely. That is why review steps belong in every use case where errors have consequences.
Are newer models more reliable?
Often yes, but not error-free. Even powerful models hallucinate, especially on rare topics, exact figures and source citations.
Does it help to ask the AI whether its answer is correct?
Only to a limited extent. A self-check can uncover errors but is not proof. Comparing against an independent source is more reliable.
Do employees need to be informed about hallucinations?
As part of AI literacy under Article 4 of the EU AI Act, the European Commission recommends that employees who use chatbots are aware of risks such as hallucinations.
Sources
- A Survey on Hallucination in Large Language Models (Huang et al., ACM TOIS) arXiv
- OWASP Top 10 for LLM Applications 2025 OWASP Foundation
- Generative AI Models – Opportunities and Risks for Industry and Authorities German Federal Office for Information Security (BSI)
- AI Literacy – Questions & Answers European Commission
Written by the Lurus Editorial Team according to our source standard. Editorial policy