AI Hallucination Explainer
An AI hallucination is an output that sounds plausible but is unsupported, inconsistent with the source material, or factually wrong. The term is most often used for generative systems that invent names, dates, quotations, citations, events, or explanations while presenting them in the same confident style as accurate information.
A language model generates text by estimating likely continuations from patterns learned during training and from the current prompt. That process does not automatically consult a database of verified facts. If the prompt is ambiguous, the needed information is missing, or several patterns compete, the model can produce a coherent continuation without reliable evidence behind it.
Hallucination rates depend on the task, model, prompt, available context, and evaluation method. Retrieval, tool use, careful instructions, constrained outputs, and abstaining when evidence is weak can reduce errors, but none guarantees truth. Even a supplied document can be misread or cited for a claim it does not support.
Users should verify consequential claims against primary sources and treat confident wording as style rather than proof. Developers can test factuality, trace claims to evidence, monitor failure patterns, and design interfaces that expose uncertainty. The practical goal is not merely fluent output, but output whose important claims can be checked.
An AI hallucination is a fluent output that is unsupported by evidence, inconsistent with the input, or factually wrong.
Key terms include factuality, grounding, unsupported claim, fabrication, attribution, retrieval, abstention, verification and uncertainty.
Generative models optimize for likely sequences, not for an internal guarantee that every statement is true. Missing context and uncommon facts make unsupported continuations more likely.
A model may also combine real details incorrectly, invent a citation-shaped reference, or follow a false assumption embedded in the question.
Grounding answers in trusted material, requiring claim-level evidence, using tools, and allowing the system to say it does not know can reduce risk.
Evaluation should test the exact use case because a model that performs well on one factual benchmark may still fail on a different domain or document set.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.