factuality

**Factuality** is the **degree to which generated output is consistent with verified knowledge and grounded source evidence** - high factuality is essential for trustworthy AI assistance in information-critical tasks. **What Is Factuality?** - **Definition**: Accuracy property measuring correspondence between model output and real-world facts. - **Evidence Basis**: Can be evaluated against trusted references, retrieved documents, or knowledge bases. - **Failure Modes**: Includes fabricated entities, incorrect relations, and outdated assertions. - **Evaluation Methods**: Human judgment, fact-checking pipelines, and entailment-based scoring. **Why Factuality Matters** - **User Trust**: Reliable factual answers are fundamental for sustained product adoption. - **Decision Safety**: Incorrect facts can produce high-cost errors in professional workflows. - **Compliance Pressure**: Regulated environments require traceable factual support. - **Brand Risk**: Frequent factual errors create reputational and legal exposure. - **System Utility**: Factual performance determines real-world value beyond language fluency. **How It Is Used in Practice** - **Grounded Generation**: Constrain responses to retrieved or curated source material. - **Citation Requirements**: Require source-backed claims for high-confidence outputs. - **Monitoring Programs**: Track factuality metrics over time and by domain segment. Factuality is **a core quality dimension for production LLM systems** - strong factual grounding, evaluation, and enforcement are required to deliver dependable answers at scale.