Hallucination (AI)
Also called confabulation.
When an AI model produces information that is false or made up, presented as if it were true.
Last updated
What it is.
A language model writes the most likely continuation of the text, not a looked-up fact. When it lacks the information, it can still produce fluent, confident text: a citation that does not exist, a wrong figure, a policy that was never written.
It cannot be removed entirely, but it can be made much rarer and easier to catch. Grounding answers in your own data with RAG, using tool calling for live facts, asking for citations and checking them, and allowing the model to say it does not know all help.
Why it matters
In business software a confident wrong answer can be worse than no answer: a wrong dosage, a wrong contract term, a wrong refund amount. Products that use AI for facts need a design that makes errors visible and checkable, plus human review where the stakes are high.
Common mistakes.
- 01
Trying to fix it with prompt wording alone instead of grounding and checks.
- 02
Hiding sources from users, so they cannot verify answers.
- 03
Not measuring it. Track the rate in evals and in production samples.
Work it out with a tool.
Related terms.
- RAG (retrieval-augmented generation)A pattern that finds relevant passages from your own data and gives them to a language model along with the question, so it answers from your content.
- Guardrails (AI)Checks placed around an AI model that block, fix or flag unsafe, off-topic or incorrect inputs and outputs.
- LLM evals (evaluations)Repeatable tests that score how well a model, prompt or AI feature does a specific task.
- Temperature (LLM setting)A setting that controls how random a language model's output is: lower is more predictable, higher is more varied.
Guides and pages
Browse the full glossary or read answers to common questions.
