Definition
Hallucination happens because LLMs predict likely text, not verified facts. They can invent citations, statistics, names, or API parameters that look correct but are wrong.
It is a core reliability risk for production AI — especially in medical, legal, or low-resource language contexts with sparse training data.
In simple terms
An LLM hallucinating is like an eloquent speaker who fills gaps in memory with convincing fiction — smooth delivery, wrong details.
Where you see it
- Legal research tools have cited non-existent cases.
- Support bots invent refund policies.
- RAG and grounding reduce but do not eliminate hallucinations.
How it works
1.Pattern completion
The model continues text that statistically fits, not text that is true.
2.Missing knowledge
Niche or recent facts may never appear in training data.
3.Mitigate
Use RAG, citations, lower temperature, and human review for high-stakes outputs.
Why it matters
- Understanding hallucination is essential before trusting LLM output in user-facing products.
Often confused
Hallucinations mean the model is broken.
They are an expected behavior of generative LLMs; engineering and UX must account for them.