Technology & AI
What is Hallucination (AI Hallucination)?
Hallucination is when the LLM invents confident but false information — inherent to the model, not a bug; mitigated with RAG, prompt and validation.
LLMs don't "know" what they know — they compute probability. When context lacks info, they fill with the "most statistically likely", even if inventing a person's name, citation or function.
Mitigation: 1) RAG (pass real info in context); 2) prompts asking for citations ("only based on text X"); 3) automated validation (regex, schema, compare to source); 4) "I don't know" as a valid response in the prompt.
Frequently asked questions
Will hallucination disappear with new models?
Decreases but doesn't vanish — it's a fundamental property of probabilistic models. Accept and protect with guardrails.
How to detect hallucination in production?
LLM-as-judge (another model validates the answer), explicit rules (regex, schema), and periodic manual sampling.