Hallucination is a large language model (LLM) giving a fluent, confident answer that is false or not supported by its sources. The model writes likely text, and likely is not the same as true.
Why it happens
A language model is trained to predict the next token. It writes a plausible continuation, and nothing checks it against the world. On subjects the model knows well, the plausible answer is usually right. On the rest, it still writes something shaped like an answer: a citation with real-sounding authors, a function the library does not have, a wrong date stated flatly.
It is more likely when:
- The question is about something rare in the training data, or newer than it.
- The prompt presupposes something false, and the model goes along.
How to reduce it
- Retrieval. Put the source text in the prompt and tell the model to answer from it. This is retrieval-augmented generation (RAG): see fine-tuning vs RAG.
- Constrained output. Require a quote or a citation for each claim, and let the model say it does not know.
- Checks on the answer. Compare the reply with its sources, or have a second model verify it, before the user sees it. See output guardrails.
These lower the rate. None brings it to zero.
With White Circle
Control analyzes every output in real time, and hallucinations (fabricated facts) are among what it covers. A reply that breaks a policy is blocked, rewritten or flagged.
