Human Evaluation: When Automation Isn't Enough

~12 min read

Human evaluation is the gold standard for subjective, high-stakes, or novel quality dimensions. Its value depends entirely on clear annotation guidelines and measuring inter-rater agreement.

Human Evaluation: When Automation Isn't Enough is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • Human evaluation is the gold standard for subjective, high-stakes, or novel qualities — and the ground truth used to calibrate automated metrics and LLM judges
  • Annotation guidelines turn 'is this good?' into a repeatable task: concrete level definitions, examples, edge-case rules, and separated dimensions
  • Inter-rater agreement (Cohen's/Fleiss' kappa) measures whether guidelines produce consistent judgments, correcting for chance agreement
  • Kappa below ~0.6 signals ambiguous guidelines or a genuinely hard task — fix and re-measure before trusting the labels
  • Use human eval strategically on a representative sample; run it as the periodic ground-truth check atop cheaper automated layers (an eval pyramid)