What is an AI hallucination?
An AI hallucination is a false or freely invented output from a language model that sounds convincing and confident but is not backed by facts.
DEFINITION
When working with a language model, hallucinations appear sooner or later. The model cites studies that do not exist, names people with wrong positions or describes historical events incorrectly — all convincingly worded and without any hint of uncertainty. This happens because an LLM does not retrieve facts but generates word sequences that sound statistically plausible. When there were no good training data for a matter or the model is computing in a grey zone between several answers, it invents. Hallucinations occur especially with very specific questions, in long conversations and under pressure to deliver an answer the model does not know. The approach is clear: verify every factually relevant statement from AI outputs, especially before reusing them. Blind trust in AI outputs is the most common mistake in AI use.
CONNECTIONS
Leadership
Teams using AI need psychological safety so employees can raise mistakes and uncertainties. Those afraid to question AI outputs pass on hallucinated information.
Agility
Agile teams using AI need a clear definition of done for AI-generated outputs. Hallucinations happen, so fact-checking must be an explicit part of acceptance criteria.
Project Management
AI hallucinations are a measurable project risk, especially when AI is used for research, analysis or reports. The risk register should explicitly track AI-specific quality risks.
KEY POINTS
- AI hallucinations are convincingly worded false answers with no factual basis.
- They arise because LLMs compute probabilities, not retrieve facts.
- Specific questions, patchy training data and conversational pressure encourage hallucinations.
- Factually relevant AI statements must always be checked.
- RAG and fine-tuning can reduce hallucinations but cannot eliminate them.
EXAMPLE
You ask a language model about a study on the effectiveness of a leadership training programme. The model names author, publisher, year and a plausible- sounding core finding. You search for the study and cannot find it. The model invented it because it saw a similar style and similar content in context and generated a convincing-sounding answer from that. So always: check linked sources before reusing them.
MISCONCEPTIONS
Does an LLM know when it is hallucinating?
No. The model has no access to meta-knowledge about its own correctness. It generates the most plausible continuation of its output, whether that is factually correct or not.
Do only cheap or small models hallucinate?
No. Very large and capable models hallucinate too. The rate drops with better models, but it does not disappear.