Hallucination
As of:
A hallucination is an output of an LLM that is plausibly phrased but factually wrong or entirely invented — fabricated references, incorrect figures or non-existent API functions, for example. The term has stuck even though it is technically off: the model is not “mistaken”; it does exactly what it was trained to do.
The cause lies in how these models work: language models generate text by predicting probable token sequences — not by looking up facts in a database. If the training data does not cover a question, or if the most probable continuation happens to be wrong, the result is fluent, convincing-sounding text without factual grounding. The linguistic confidence of an answer is therefore no indicator of its correctness.
In practice, hallucinations are the most important risk in production use of generative AI, especially in sensitive domains such as law, medicine or finance. Proven countermeasures: connect vetted sources via RAG, require verifiable citations, validate outputs automatically and have humans review critical results. A common misconception is that hallucinations can be eliminated entirely — as things stand, they are an inherent property of the technology. Processes should therefore be designed to expect erroneous outputs.