An AI hallucination is an output from a language model that sounds plausible but is false, irrelevant or invented. The word also has a medical meaning, a perception without an external stimulus. This entry covers the AI meaning.
What counts as a hallucination?
IBM describes hallucinations as outputs that sound plausible but are factually wrong, irrelevant or entirely fabricated. The NIST profile for generative AI uses the term confabulation for confidently stated but false content and notes that people also call it hallucination. Typical cases are invented sources, wrong figures and made-up product details.
A well-known example is a legal brief in which a lawyer cited court cases that ChatGPT had invented. Both IBM and The Conversation report the case.
Why do models hallucinate?
A language model predicts likely continuations of text. It does not look facts up by default. IBM and The Conversation list several contributing factors, among them gaps or errors in the training data, pattern matching without grounding in verified sources, and ambiguous questions.
How can the risk be reduced?
Common measures are retrieval-augmented generation (RAG), where the system fetches relevant documents before answering, clear instructions in the prompt, and human review of important outputs. Anthropic says RAG is meant to support answers with evidence. Retrieval lowers the risk, but it does not remove it, because the quality of the retrieved sources still matters.
What does it mean for SEO and GEO?
AI search answers can contain wrong statements about companies, products or prices. As an assessment, it makes sense to check regularly what AI assistants say about your own brand. Pages that are current, unambiguous and easy to retrieve give these systems better sources to draw from.
Example
Asked for the price of a software product, a model names a specific monthly fee that appears on no page of the vendor. The tone is confident, the number is made up.
FAQ
Does AI still hallucinate? Yes. NIST lists confabulation as a risk of generative AI, and vendors describe mitigation, not elimination.
Can you stop an AI from hallucinating? Not completely. Retrieval from reliable sources, clear prompts and review reduce the rate of errors.
Sources
Last updated: