What is Hallucination (AI)?
When an AI model generates information that is fluent and confident-sounding but factually incorrect, fabricated, or unsupported by its training data. Hallucinations can range from subtle (wrong dates, invented citations) to egregious (entirely fictional events presented as fact). The term is debated (some prefer "confabulation") but it is the industry standard.
My perspective
In practice
Every AI output should be treated as a draft, not a fact. Hallucination risk increases with obscure topics, numerical claims, and questions about specific people or events. Mitigation strategies include grounding models in retrieved documents (RAG), asking for sources and verifying them, and using models with lower temperature settings for factual tasks. Trust but verify. Always.