Hallucinations are confident, plausible-sounding, incorrect outputs. They are not bugs — they are the model's default behavior when its training distribution does not cover the query. Understanding causes, categories, and mitigations is the difference between deploying a demo and deploying a system.
- 01Distinguish extrinsic from intrinsic hallucinations.
- 02Name three mechanistic causes.
- 03State the three primary mitigations and their limits.
The taxonomy
Extrinsic hallucination: the output contradicts the input (misquotes a document, invents a citation). Intrinsic hallucination: the output contradicts world facts. Both are real failure modes with different mitigations.
The causes
1. Training on incorrect data. 2. Optimization pressure toward fluent, confident-sounding responses. 3. RLHF may reward-hack confidence — models learn users prefer confident-sounding answers.
The mitigations
Retrieval grounding (RAG), tool use (calculator, search), and explicit uncertainty prompting reduce hallucination but do not eliminate it. A production system needs verification, not just prevention.
- Hallucination is the default; correctness is engineered on top.
- Grounding + verification is the only durable mitigation.