Hallucination.

Hallucination in AI refers to when a language model generates output that is factually incorrect, fabricated, or unsupported — yet delivered with the same confident, fluent tone as accurate content. The term is borrowed from psychology, where hallucination means perceiving things not present in reality. AI hallucination is a fundamental limitation of current LLMs, not a bug that will be fully eliminated.

Why Models Hallucinate

LLMs predict the most statistically probable next token given the input — they don't 'know' facts in a human sense; they pattern-match. When asked about something outside their training distribution, or when they lack relevant information, they generate plausible-sounding but incorrect output rather than admitting uncertainty. Hallucination rates increase on rare topics, specific statistics, and recent events.

Hallucination in AI Overviews

Google's AI Overviews have made hallucination high-profile — incorrect AI-generated information appearing at the top of SERPs can spread misinformation at scale. Google uses grounding (RAG) to reduce hallucination, but errors persist. Brands monitoring AI Overviews for factual accuracy about themselves is an emerging reputation management practice.

Mitigation Strategies

Hallucination mitigation includes: RAG (providing the model with retrieved source documents to ground responses), chain-of-thought prompting (reasoning out loud reduces confident errors), temperature reduction (lower randomness reduces creative fabrication), output verification steps (a second model checks the first's output), and human review for high-stakes content.