Think of an artificial intelligence as a student who has read an enormous number of books but did not always understand everything perfectly. Ask him about something he has no solid information on and he will not try to admit it. Instead, on the basis of learned patterns and turns of phrase, he confidently “invents” an answer that sounds highly convincing yet is factually wrong. This tendency of an AI to generate untrue, invented or misleading information not backed by real data is called hallucination. It is as if the AI told you with total certainty that the capital of Australia is Sydney.
Hallucination in the context of artificial intelligence, especially in large language models (LLMs), denotes the generation of outputs that are factually incorrect, misleading or unsupported by the source data, even though they are syntactically and grammatically coherent and appear credible. The phenomenon is not the result of conscious intent but a systemic property of current generative models. It arises for several reasons, among them noise and inaccuracy in the training data, the model’s limited understanding of the real world, and an excessive drive to generate fluent, probable text at the expense of factual correctness.
When generating an answer, the model does not attempt to verify facts against an external database; it assembles a sequence of words on the basis of the statistical probabilities it absorbed during training. If a query is ambiguous or touches an area where the model has thin data coverage, it may extrapolate in a way that produces a plausible-sounding but fabricated answer. Detecting and mitigating hallucinations is one of the active areas of research, using techniques such as RAG (Retrieval-Augmented Generation) to ground the model in verifiable information.