Hallucinations: Why LLMs Make Things Up

The big idea: A large language model (LLM) can hand you a sentence that is smooth, confident, grammatically perfect — and completely false. In AI, we call this a hallucination: the model states something that is not true as if it were a solid fact. It is not lying on purpose, and it is not broken. It is doing exactly what it was built to do — produce text that sounds right. Whether that text is actually right is a separate question the model never truly checked.

This lesson teaches you to expect hallucinations, spot them, and defend against them with a simple habit: trust, but verify.

• What a hallucination is, in plain language: confident, fluent, and wrong.
• The root cause: an LLM predicts plausible text; it does not look facts up in a truth database.
• The classic danger zones where hallucinations happen most (citations, statistics, laws, medical/financial advice, biographies, recent events).
• Real cautionary tales — including lawyers punished for citing court cases an AI invented.
• How to detect a likely hallucination before it burns you.
• A repeatable verification workflow so you never publish unchecked AI claims.
• Why grounding (RAG and web-connected modes) reduces hallucinations but never fully removes them.

A hallucination is any output from a language model that is presented as fact but is false, unsupported, or invented. The tricky part is how it is wrong. The model does not stammer, hedge, or leave a blank. It fills the gap with something that looks perfectly reasonable — a plausible name, a plausible date, a plausible quote, a plausible web link that leads nowhere.