Factuality

Factuality describes whether AI-generated claims correspond to reality or to an authoritative source of truth. It is especially important for language models that can produce fluent statements without knowing whether they are correct. Evaluation may use reference answers, retrieved evidence, citation checking, structured databases, expert review, or specialized fact-verification models. No single score captures every domain, and creative or hypothetical tasks do not always require literal factual output. Factuality should also be distinguished from groundedness: a response can accurately reflect a supplied document even if that document is wrong. High-stakes systems need current sources, clear uncertainty, traceable evidence, and human review in addition to automated metrics.