---
title: "Hallucination Rate"
slug: "hallucination-rate"
category: "glossary"
tags: ["glossary", "ai-agents", "llm", "evaluation", "reliability", "metrics"]
status: "stable"
last_updated: 2026-05-14
summary: "Hallucination rate is the proportion of model outputs with factually incorrect or fabricated content, measured against ground truth or human annotation."
related:
  [
    "[[glossary/hallucination]]",
    "[[glossary/ground-truth]]",
    "[[glossary/eval-set]]",
    "[[glossary/rag]]",
    "[[glossary/evaluation-harness]]",
    "[[glossary/llm-as-judge]]",
  ]
---

## Overview

This page is the atomic definition. The underlying phenomenon is defined at [[glossary/hallucination]].

## Definition

Hallucination rate is a quantitative metric that measures the frequency of factually incorrect or unsupported statements in LLM outputs over a defined evaluation set. It is computed as: (number of outputs with at least one hallucination) / (total outputs evaluated). The metric requires a reference: either a [[glossary/ground-truth]] dataset with verified correct answers, or human annotators rating each claim. Automated evaluation using [[glossary/llm-as-judge]] systems (one LLM critiquing another) scales better than human annotation but introduces its own biases. Hallucination rate is task-specific: a model may have near-zero hallucination rate on common knowledge questions and 30% on niche domain questions. [[glossary/rag]] reduces hallucination rate on retrievable facts but can introduce grounding errors when retrieved context is irrelevant or contradictory. Tracking hallucination rate over model versions and prompt changes is a prerequisite for safe production LLM deployment.

## When it applies

Measure hallucination rate before deploying any LLM system where incorrect facts cause harm: medical, legal, financial, or safety-critical applications. Monitor it in production using sampling and spot-check annotation. Set an acceptable threshold (e.g., under 2%) and block model or prompt upgrades that exceed it.

## Example

An [[glossary/eval-set]] of 500 drug interaction questions is evaluated against a pharmacist-verified answer key. The model answers 487 correctly and includes one or more incorrect claims in 13 answers. Hallucination rate = 13/500 = 2.6%. The threshold is 2%; the model fails and is not promoted to production.

## Related concepts

- [[glossary/hallucination]] - the underlying phenomenon; hallucination rate quantifies it.
- [[glossary/ground-truth]] - the verified reference required to detect hallucinations.
- [[glossary/eval-set]] - the test set over which hallucination rate is measured.
- [[glossary/rag]] - retrieval reduces hallucination rate on knowledge-intensive tasks.
- [[glossary/llm-as-judge]] - automated scoring scales hallucination rate measurement.

## Citing this term

> See [[glossary/hallucination-rate|Hallucination Rate]] (llmbestpractices.com/glossary/hallucination-rate).

## Related

- [[glossary/hallucination]]
- [[glossary/ground-truth]]
- [[glossary/eval-set]]
- [[glossary/rag]]
- [[glossary/evaluation-harness]]
- [[glossary/llm-as-judge]]