Fairness Metrics
Definition
Fairness metrics formalize different conceptions of equitable treatment into measurable quantities. Key metrics include: (1) Statistical parity difference—the difference in positive prediction rates between groups (target: < 0.05); (2) Disparate impact ratio—the ratio of positive rates between groups (target: > 0.8, the '80% rule' from EEOC guidelines); (3) Equal opportunity difference—the difference in true positive rates between groups (target: < 0.05); (4) Equalized odds difference—the maximum of true positive and false positive rate differences (target: < 0.05); (5) Average odds difference—average of TPR and FPR differences; (6) Calibration by group—whether predicted probabilities match observed frequencies equally across groups. Different metrics make different normative assumptions about what fairness means.
Why It Matters
Fairness metrics are the operational tool that converts abstract fairness principles into measurable audit criteria. Without metrics, bias evaluation is subjective and incomplete; with metrics, teams can set acceptance thresholds, track fairness over time, compare models by fairness profile, and demonstrate compliance with regulatory requirements. The EU AI Act, Equal Credit Opportunity Act, and Title VII of the Civil Rights Act all require that AI systems making consequential decisions don't discriminate—fairness metrics are how you measure and document compliance. They are also debugging tools: a high average accuracy but poor equalized odds immediately tells you where to investigate.
How It Works
Fairness metric computation requires: (1) a test dataset with both model predictions and demographic labels for the relevant protected attributes; (2) binary or multi-class outcome labels; (3) computation of confusion matrix components (TP, FP, TN, FN) per demographic group; (4) calculation of each fairness metric from these components; (5) comparison against defined thresholds. Python toolkits (Fairlearn, AI Fairness 360, Aequitas) automate this computation with visualization and reporting. The choice of reference group matters—metrics compare an unprivileged group to a privileged group, so the comparison baseline must be defined explicitly.
Fairness Metrics Dashboard
Demographic Parity Difference
0.08 (threshold: ≤0.10)
PASSEqualized Odds Difference
0.14 (threshold: ≤0.10)
FAILPredictive Parity Difference
0.06 (threshold: ≤0.10)
PASSIndividual Fairness Score
0.91 (threshold: ≥0.85)
PASSReal-World Example
A consumer credit company computed 6 fairness metrics for their new credit approval model before regulatory submission. Five of six metrics passed their thresholds. The failing metric was false positive rate parity—the model incorrectly denied credit to 31% of actually-qualified Hispanic applicants vs. 18% of qualified white applicants. This asymmetric false positive rate meant qualified minority applicants were more likely to be wrongly denied credit, a disparate harm not captured by aggregate accuracy metrics. The model required architecture revision to address the false positive disparity before regulatory approval—a process that took 8 weeks but prevented a CFPB enforcement action.
Common Mistakes
- ✕Reporting only aggregate fairness metrics without disaggregated group-level results—averages can hide severe disparities affecting small groups
- ✕Choosing fairness metrics without considering what type of error is most harmful in context—a false positive in criminal justice (detaining innocent people) has different stakes than in lending
- ✕Treating compliance with fairness thresholds as a guarantee of fairness—metrics measure specific defined fairness criteria; real-world fairness requires broader context
Related Terms
AI Bias
AI bias is the systematic tendency of AI models to produce unfair outcomes for certain groups—arising from skewed training data, biased features, or flawed objective functions—leading to discriminatory predictions or decisions.
Algorithmic Fairness
Algorithmic fairness defines formal mathematical criteria for measuring and achieving equitable treatment across demographic groups in AI decision systems—including demographic parity, equalized odds, and individual fairness.
Disparate Impact
Disparate impact occurs when an AI system produces significantly different outcomes for different demographic groups—even without explicitly using protected attributes—creating legal liability under anti-discrimination law regardless of intent.
Responsible AI
Responsible AI is a framework of organizational practices and principles—encompassing fairness, transparency, privacy, safety, and accountability—that guide how teams build and deploy AI systems that are trustworthy and beneficial.
Explainability
Explainability provides human-understandable reasons for why an AI system produced a specific output—enabling users, operators, and regulators to understand, audit, and trust AI decisions rather than treating the model as an inscrutable black box.
Ready to build your AI chatbot?
Put these concepts into practice with 99helpers — no code required.
Start free trial →