- Automated comparison of generated text against a knowledge base?
- Measuring the model’s confidence scores for each generated statement?
- Human expert verification combined with factual knowledge-base checking?
- Counting the number of citations included in the model’s response?
- Define a representative, diverse test set of prompts and enumerate the specific factual claims or expected answers for each prompt. Include edge cases and ambiguous contexts.
- Recruit multiple domain experts to independently annotate model outputs for factuality and error type (e.g., fabrication, misattribution, omission, partial truth). Require annotators to record brief justifications for their labels to support adjudication.
- Run automated fact checks against one or more curated knowledge bases or canonical sources to flag claims that are clearly verifiable or falsifiable.
- Create a decision rule to combine expert annotations with automated flags into a final label for each claim (for example: factual / partially factual / hallucinated).
- Quantify annotation reliability with inter-annotator agreement measures such as Cohen’s kappa and Krippendorff’s alpha.
- Report both aggregate metrics and representative qualitative examples. Include edge-case analyses to clarify limitations and interpretability.

- Automated comparison against knowledge bases is useful for high-precision checks but can miss context-dependent or nuanced errors and is limited by the coverage and freshness of the knowledge base.
- Model confidence scores (e.g., logit-based or probability measures) do not consistently correlate with factual accuracy and can give a false sense of reliability.
- Counting citations is insufficient because models can fabricate references or include irrelevant/incorrect citations; citation presence does not guarantee factuality.
Best practice: adopt a hybrid evaluation—structured human annotation plus targeted automated checks. Publish both quantitative metrics and curated qualitative examples so readers can assess strengths, failure modes, and real-world applicability.
- Cohen’s kappa — measure of inter-rater reliability
- Krippendorff’s alpha — flexible agreement coefficient for multiple raters and data types
- For practical toolchains, combine annotation platforms (for expert labeling) with programmable fact-checking pipelines that query curated databases or APIs.
- When publishing evaluations, include dataset examples, annotation guidelines, and adjudication rules so results are reproducible and actionable.