What is AI Hallucination? - Definition & Meaning
Learn what AI hallucination is, why LLMs make up facts, and which techniques to use to reduce hallucinations in production.
Definition
AI hallucination is the phenomenon where an LLM produces plausible-sounding but incorrect or fabricated information — not based on input or factual knowledge. It is one of the biggest challenges for reliable AI.
Technical explanation
Causes: models generate probabilistically; they choose tokens that are "likely", not necessarily correct. Confabulation: the model fills gaps with what seems consistent. There are fabrications (fully made up), extrapolations (taken too far), and attribution errors (wrong source). Mitigation: RAG (ground in documents), cite sources, reduce open-ended generation, guardrails, fact-checking layers, smaller/scoped prompts. Newer models hallucinate less but the problem remains.
How AVARC Solutions applies this
AVARC Solutions designs AI systems with hallucination mitigation: RAG where possible, structured output, guardrails for sensitive claims, and human review for high-stakes decisions. We test on edge cases where models tend to hallucinate.
Practical examples
- A chatbot mentioning a non-existent product because it follows the pattern of a product name.
- A summarizer adding facts not present in the source text.
- A support agent citing an outdated policy that no longer applies.
Related terms
Frequently asked questions
Related articles
What is RAG (Retrieval Augmented Generation)? - Definition & Meaning
Learn what RAG is, how it combines LLMs with external knowledge sources for accurate and up-to-date answers, and why it is essential for enterprise AI.
What are AI Guardrails? - Definition & Meaning
Learn what AI guardrails are, how they filter and constrain output for safety and compliance, and which tools to use for production AI.
What is Contextual Compression? - Definition & Meaning
Learn what contextual compression is, how retrieved documents are compressed based on the query, and why it makes RAG more efficient and effective.
RAG Application Template - Retrieval Augmented Generation Setup
Download our RAG application template for knowledge base chatbots and Q&A systems. Includes chunking, embeddings, vector database, and prompt design.