Hallucinations & Grounding
Theory
Hallucination is output that sounds correct but is factually wrong. The model has no built-in truth-checker — pretraining rewards predicting plausible tokens, not accurate ones.
| Type | What it contradicts |
|---|---|
| Intrinsic | Something stated in the input or context |
| Extrinsic | External reality — fabricated fact, citation, or statistic |
Ungrounded — fabricated
Q: Who cited Einstein in the 1905 relativity paper?
A: Poincaré and Lorentz are cited in footnote 12.
Confident, fluent, false. The footnote does not exist.
Grounded with retrieved context
Context: [1905 paper text — no footnotes]
A: The paper contains no citation footnotes.
Answer drawn from supplied facts, not pattern memory.
CoT is not a fix: a fluent reasoning chain can still end at a wrong conclusion. Reach for retrieval grounding (RAG), source-citation prompts, lower temperature, and constrained decoding. None guarantees accuracy. Act 3 covers retrieval in depth.