Factuality

Full definition

Factuality is measured with benchmarks such as TruthfulQA, FActScore, and SimpleQA, plus retrieval-grounded checks where each claim must be supported by a cited source. The opposite of factuality is hallucination. Frontier LLMs hallucinate at rates between 3% and 27% on long-form generation depending on domain — making factuality a first-class governance metric.

Why it matters

For regulated industries, factuality is the gating constraint on LLM deployment. A model that is 5% wrong on medical or legal facts is unshippable regardless of how fluent it is. Buyers should demand factuality numbers, not just leaderboard scores.

Example

A bank's internal copilot is benchmarked on a 2,000-question regulatory-knowledge dataset; factuality below 95% blocks promotion to production until RAG grounding is added.

Related terms

• Hallucination
  An AI hallucination is when a language or generative model produces content that is plausible-sounding but factually incorrect, fabricated, or unfaithful to its sources — the single biggest failure mode of LLMs in regulated and high-stakes domains.
• Retrieval-Augmented Generation (RAG)
  Retrieval-Augmented Generation (RAG) is a technique where a language model retrieves relevant documents from an external knowledge base before generating a response, reducing hallucination by 30-60% and enabling citation of sources.
• Model Evaluation
  Model evaluation is the systematic measurement of an AI model's performance, safety, and behavior across representative tasks — using fixed benchmarks, golden datasets, human ratings, and LLM-as-judge methods — both before and after deployment.
• AI Guardrails
  AI guardrails are the layered controls — input filters, output classifiers, policy engines, schema validation, and structured generation — that constrain an AI system to safe, on-policy, and on-task behavior.