Hallucination (AI)
When an AI model generates plausible-sounding but factually incorrect information. A fundamental limitation of LLMs.
An AI hallucination is a confidently stated piece of information that's factually wrong. LLMs hallucinate because they generate text by predicting likely sequences, not by retrieving verified facts. When the model doesn't know something, it often produces something that sounds right but isn't.
Examples: inventing book titles, fabricating quotes, citing papers that don't exist, generating incorrect math, making up product features.
Mitigation strategies include: using RAG to ground responses in real documents, asking the model to cite sources (like Perplexity does), training with RLHF to discourage confident wrong answers, and post-hoc fact-checking. No technique fully eliminates hallucinations in current LLMs. Critical workflows should always verify AI outputs before acting on them.