AI Hallucinations Explainer
An AI hallucination is an answer that sounds plausible but contains false, invented, unsupported or internally inconsistent information. The National Institute of Standards and Technology uses the term “confabulation” for this behavior. It can appear as a made-up citation, an incorrect date, a fabricated quotation or a confident explanation that does not follow from the evidence.
Generative models do not consult a built-in ledger of verified facts each time they respond. A language model predicts a sequence of tokens from patterns learned during training and from the context it receives. That process can reproduce accurate information, but it can also complete a pattern with details that fit linguistically rather than factually—especially when a prompt is ambiguous, the needed information is missing or the model is pushed beyond its knowledge.
Hallucinations matter most when people mistake fluency for reliability. A wrong answer about a movie plot may be harmless; a false medical summary, legal citation or financial claim can cause real damage. Retrieval, citations, tools and better prompting can reduce errors, but none of them makes every output true. Retrieved material can itself be irrelevant or outdated, and a model can misread a good source.
Safer use combines model design with human judgment. Ask for sources, open those sources, compare consequential claims with primary records and give the system permission to say that evidence is insufficient. Developers can test factual accuracy, calibrate uncertainty, constrain outputs and route high-stakes decisions to qualified people. The practical rule is simple: treat generated text as a draft to verify, not as proof.
An AI hallucination is a confident output that is false, unsupported or inconsistent. Related terms include confabulation, factuality, grounding, calibration, uncertainty, citation and provenance.
NIST’s Generative AI Profile defines confabulation and explains its risks. Google Research examines confident errors and the role of faithful uncertainty.
A language model generates likely token sequences. It may have incomplete knowledge, receive too little context or favor a fluent continuation even when no well-supported answer is available.
Sampling settings, vague questions, conflicting context and requests for obscure details can increase error risk. A model may also attach invented citations or reasoning to a wrong conclusion.
Ground the task in reliable documents, require claim-level citations, test whether citations support the claim and compare high-stakes answers with primary records.
Risk can be reduced, not eliminated. Systems should express uncertainty, decline when evidence is insufficient and keep qualified humans responsible for consequential decisions.
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.