Hallucination rate is a quantitative metric expressing how frequently a given model produces factually incorrect, fabricated, or unverifiable content across a defined evaluation set. It is typically reported as a percentage — for instance, a hallucination rate of 8% on a benchmark means that roughly 1 in 12 responses contained at least one factual error. Because "hallucination" is not a single uniform failure mode, reported rates depend heavily on the benchmark methodology, the subject domain, the prompt style, and whether evaluation is done by human annotators, automated classifiers, or a secondary LLM acting as judge.

Think of it like a defect rate in manufacturing: it tells you the proportion of outputs that don't meet the quality standard, but it doesn't tell you why individual defects occur or how severe each one is. A 2% hallucination rate on factual Q&A does not mean 98% of individual facts in those answers are correct — it means 98% of full responses were judged acceptable under the evaluation rubric.

Measuring hallucination rate concretely requires a ground-truth dataset. Researchers often use retrieval-augmented benchmarks where the model is given source documents and asked questions whose answers appear verbatim in the text; any deviation is a candidate hallucination. Example: the FRAMES benchmark presents multi-hop factual questions; a model scoring 72% is hallucinating on roughly 28% of responses. Proprietary evaluations by AI labs compare model outputs against curated fact stores annotated by domain experts.

Hallucination rate is one of several reliability metrics alongside Perplexity and benchmark scores; it is distinct from but related to calibration (how well a model's expressed confidence matches its accuracy). Techniques like RAG, Fine-tuning on high-quality corpora, and RLHF with factual reward signals all aim to reduce it.

You will encounter hallucination rate when comparing LLMs for deployment in medical, legal, or financial applications, and in model cards and evaluation reports published alongside major model releases.