Huang et al. (2023) published a comprehensive survey on the hallucination problem in large language models — the tendency of AI to generate plausible but factually incorrect information:
Types of Hallucination
- Factual hallucination: Generating false facts with high confidence
- Faithfulness hallucination: Contradicting the source material or prior context
- Input-conflicting: Generating outputs that diverge from the provided input
Why It Matters
- LLMs hallucinate in 3-27% of responses depending on the task and model
- Hallucinations are fluent and confident, making them difficult for users to detect
- In high-stakes domains (healthcare, law, finance), even rare hallucinations can have severe consequences
- Users who experience automation bias are particularly vulnerable to accepting hallucinated content as fact
Source
Huang, L. et al. (2023). A Survey on Hallucination in Large Language Models. arXiv:2311.05232.