What Is Hallucination?
LLM hallucination occurs when the model generates confident, fluent, plausible-sounding text that is factually wrong. It is not random noise — the model truly "believes" its output. This makes hallucination insidious: wrong information delivered with the same tone as correct information.
Why Do LLMs Hallucinate?
- Training objective: Models are trained to predict the next token, not to be accurate. Fluency and accuracy are correlated but not identical.
- Knowledge boundaries: Models do not know what they do not know. There is no internal "I am not sure" signal in vanilla autoregressive generation.
- Sycophancy: RLHF-trained models are rewarded for responses humans rate highly — confident, fluent answers score better than "I don't know."
- Compression: LLMs compress world knowledge into a fixed parameter count. Rare or complex facts get compressed imperfectly.
Types of Hallucination
- Factual hallucination: Wrong facts — "Einstein won the Nobel Prize for relativity" (it was for the photoelectric effect).
- Citation hallucination: Fabricated papers, authors, URLs that do not exist.
- Instruction hallucination: The model ignores or misunderstands its instructions.
- Reasoning hallucination: Correct premises, wrong conclusion due to flawed reasoning steps.
Prevention Strategies
RAG: Ground Responses in Documents
The most effective intervention. If the model can only answer from retrieved documents, it cannot hallucinate facts outside those documents.
Self-Consistency
Generate the same answer multiple times with temperature > 0, then take the majority vote. Hallucinated facts rarely appear consistently across multiple samples.
Verification Prompting
prompt = """Answer the question. Then review your answer:
1. Is every factual claim in your answer verified?
2. If you are uncertain about anything, say so explicitly.
Question: {question}"""Adding "If you are not sure, say 'I don't know'" to your system prompt reduces hallucination rates by 20–35% across most models.
Automated Hallucination Detection
Tools like Ragas, TruLens, and Arize Phoenix can score generated answers against retrieved context, flagging responses where the model's claims are unsupported by the source documents.
