Hallucination is the term used when an AI text-generation system produces output that sounds confident and fluent but is factually wrong, fabricated, or not supported by the information it was given. A model might invent a citation, state a false date, or describe a product feature that does not exist - all in perfectly natural prose. The fluency is what makes hallucination dangerous: the text gives no signal that it is unreliable.
The most cited treatment of the problem is the survey “Survey of Hallucination in Natural Language Generation” by Ziwei Ji and colleagues, posted in 2022. The authors describe how deep-learning text generators are “prone to hallucinate unintended text, which degrades the system performance and fails to meet user expectations in many real-world scenarios.” Rather than offering a single formula, the survey catalogs how the problem shows up across summarization, dialogue, machine translation, and other tasks, and reviews the methods proposed to measure and reduce it.
A useful distinction the literature draws is between text that contradicts the source material a model was given (sometimes called intrinsic hallucination) and text that simply cannot be verified from any source (extrinsic). Both stem from the same root: these models predict likely sequences of words, and a plausible-sounding falsehood can be statistically likely even when it is untrue.
Why business readers should care: hallucination is the single biggest reliability risk in deploying language models for anything where accuracy matters - legal, financial, medical, or customer-facing work. It is also why techniques like retrieval-augmented generation (grounding answers in a trusted document set) and human review of high-stakes outputs are not optional extras but core parts of a responsible deployment.