What is AI Hallucination?

Language models generate text by predicting the most likely next token given everything that came before. They don't look things up — they pattern-match against their training data. When asked about something outside their training, in a domain with sparse training coverage, or requiring precise factual recall, they don't say 'I don't know' — they generate fluent, confident-sounding text that follows the statistical patterns of correct answers without actually being correct. This is hallucination. Hallucinations range from subtle (slightly wrong dates or statistics) to egregious (fully fabricated court cases, invented academic papers with plausible-sounding authors and journals). The danger is that hallucinated content is typically indistinguishable in tone and fluency from accurate content — the model doesn't signal its uncertainty. Hallucination is not a bug that can be patched — it's a structural property of how autoregressive language models work. Mitigation strategies include retrieval-augmented generation (grounding responses in retrieved documents), citation requirements (asking the model to cite sources, then verifying them), temperature reduction (lower temperature produces more conservative outputs), and output verification pipelines. Hallucination rates vary significantly by domain, question type, and model — understanding this variance is essential for deciding where LLMs can and cannot be safely deployed.