An LLM hallucinates when it states something false with total confidence. It isn't lying or malfunctioning — it's doing exactly what it was trained to do: predict the most plausible next token, not the most true one.
Why it happens
- A model is a next-token predictor. "Plausible" and "true" usually coincide — but not always.
- It has no built-in fact-checker and no awareness of the boundary of its own knowledge.
- Pressed for a citation, name, or number it doesn't have, the most plausible continuation is a confident, well-formatted fabrication.
- Training rewards fluent, helpful answers — sometimes over admitting "I don't know."
It's the Chinese Room in practice: fluent symbol manipulation with no grounding in reality.
Reducing it
| Tactic | Effect |
|---|---|
| RAG | Ground answers in retrieved sources |
| Lower temperature | Less creative drift |
| Ask for citations / "say if unsure" | Surfaces uncertainty |
| Verify with tools | Don't trust facts/math blindly |
Treat an LLM as a brilliant, confident intern who never says "I'm not sure" — verify anything that matters.
Related: The Chinese Room · RAG with Local Embeddings · The Turing Test