Hallucination
When an AI model generates confident-sounding but factually incorrect or fabricated information.
Hallucination refers to the phenomenon where an AI model — particularly a large language model — generates text that is factually wrong, contradictory, or completely fabricated, while presenting it with total confidence. The model doesn't "know" it's wrong; it's simply predicting plausible-sounding tokens without grounding in truth.
Common hallucinations include: made-up citations, incorrect dates or statistics, invented quotes from real people, nonexistent software documentation, and false claims about current events. This is a fundamental limitation of the autoregressive generation process — models optimize for fluency, not factuality.
Mitigation Strategies
- RAG — Retrieval-augmented generation grounds responses in retrieved documents
- Grounding — connecting outputs to verified data sources
- Confidence signals — asking the model to express uncertainty
- Human review — always verify high-stakes AI-generated claims
Hallucination rates vary significantly by model and task type. Factual retrieval tasks (dates, names, specific numbers) see higher rates than creative or generative tasks. Newer reasoning models and retrieval-augmented systems have reduced hallucination rates, but the problem is not fully solved. Always verify AI-generated factual claims with primary sources.