AI Hallucination
AI Hallucination is when an LLM generates plausible-sounding but false or unfounded information, including making up citations, facts, or source attributions. It is a fundamental limitation of how LLMs generate text.
Definition
LLMs predict the next token based on patterns in training data, not based on truth verification. They can confidently produce convincing-sounding text that contradicts reality, attributes false facts to people, or cites sources that do not exist. Hallucinations are harder to detect than factual errors because they are fluent and contextually consistent.
In AI search contexts, hallucinations manifest as misquoted facts, wrong attributions, or invented statistics appearing in answers. These spread misinformation while damaging trust in both the answer and the cited sources. Users see fake citations to your brand, or worse, your brand is associated with incorrect information the LLM generated.
Why it matters for AI visibility
Hallucinations create AI brand safety risks. Your brand might be cited for claims you never made, or cited alongside false information that undermines your credibility. Understanding hallucination risk helps you decide whether and how to optimize for AI search, and what safeguards to implement in proprietary answer engines.
Related terms
Hallucination Mitigation
Hallucination Mitigation refers to techniques that reduce false outputs in LLM responses, including grounding in retrieved sources, fact-checking, and training adjustments. No method fully eliminates hallucination.
AIRetrieval-Augmented Generation (RAG)
RAG is the technique of retrieving relevant documents at answer time, then using them to ground an LLM's response. It enables AI search engines to cite sources while generating answers.
GEOAI Grounding
AI Grounding is the process of constraining LLM generation to reference and cite specific source documents, reducing hallucination and ensuring generated responses are anchored to retrievable content.
AIKnowledge Cutoff
Knowledge Cutoff is the date up to which an LLM was trained on text data. Information or events after this date are unknown to the model unless provided through RAG.