Hallucination

Property evaluation published

Also known as: Confabulation, Fabrication, Making Things Up

Definition

When an LLM generates content that is factually incorrect, nonsensical, or unfaithful to provided source material, presented confidently as if true. Hallucinations are a fundamental limitation of LLMs—they generate plausible- sounding text without grounded verification of truth. The term emphasizes that the model isn't "lying" but producing outputs disconnected from reality.

What this is NOT

  • Not intentional deception (models don't have intent)
  • Not bugs in the code (hallucination is inherent to generative models)
  • Not all incorrect outputs (must be presented confidently as fact)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

Model outputs that contain fabricated facts, invented citations, false claims, or information that contradicts the provided context. A major reliability challenge for LLM applications.

Sources: Hallucination in LLMs surveys, RAG hallucination research, Model evaluation literature

academic-nlp

Generated text that is unfaithful to the source material (extrinsic hallucination) or contains information not verifiable from any source (intrinsic hallucination).

Sources: Hallucination taxonomy papers, Faithfulness evaluation research

Examples

  • Model citing a paper that doesn't exist
  • Model stating incorrect dates for historical events
  • Model contradicting the retrieved context in RAG
  • Model confidently describing features a product doesn't have

Counterexamples

Things that might seem like Hallucination but are not:

  • Model saying 'I don't know' (not hallucination, appropriate uncertainty)
  • Model making a reasoning error in math (error, not hallucination)
  • Model generating fiction when asked to (intentional, not hallucination)

Relations

See Also

Implementations

Tools and frameworks that implement this concept: