What is Hallucination?
When an LLM generates confident but false content — made-up API signatures, fake citations, invented file paths. RAG, tool use (verify by running), and "read the source" patterns all reduce but don't eliminate hallucinations.
What it is
A hallucination is text that is factually wrong or inconsistent with the context it was given, produced with the same fluency as a correct answer. Anthropic notes that even the most advanced models sometimes do this, which can undermine the reliability of anything built on them.
What reduces it
Anthropic's guidance gives three basic strategies: explicitly let the model say it does not know; for long documents, ask it to extract word-for-word quotes first and answer from those; and make answers auditable by having it cite a supporting quote for each claim and retract any claim it cannot support. Advanced options include reviewing the model's reasoning when an answer looks wrong, running the same prompt several times and comparing for inconsistencies, asking follow-ups that verify earlier statements, and telling the model to use only the documents you provided.
What it does not do
These techniques reduce hallucinations; they do not eliminate them. For anything high-stakes, validate the output yourself. Agents raise the stakes because a fabricated detail can become an action, such as a wrong command, a made-up file path or an email sent with an invented fact.
Example
Asked to summarise a report, a model invents a statistic that sounds right. Telling it to quote the sentence that supports each figure, and to drop any figure it cannot quote, removes the invented number or exposes it.
Sources
← Back to the full AI agent glossary.