Bias Metric

What is Bias Metric?

A bias metric quantifies unwanted political, gender, social, or cultural skew in model outputs. It is especially important for decision systems such as loan approval or recruitment, where biased language can translate into unfair outcomes.

Bias is context-dependent: phrases that read neutral in one culture may be charged in another, which is why few-shot spot checks alone rarely give a reliable bias score.

What teams usually do about it

• Define protected attributes and disallowed stereotypes for the domain
• Evaluate across locales and demographic slices, not a single prompt set
• Track bias metrics alongside accuracy so fairness does not regress silently

Related Giskard articles

• Practical guide to LLM hallucinations and misinformation
• Announcing Giskard OSS v3

Measure bias with Giskard — combine bias metrics with scenario-based evals so fairness regressions surface before high-stakes decisions. Explore Giskard OSS evaluation checks.

Authority: NIST AI Risk Management Framework.