Description

Fairness metrics are primitives that quantify algorithmic bias across demographic groups through standardized mathematical measures. Rather than claiming systems are "fair" without evidence, fairness metrics provide concrete numerical measures of disparities: statistical parity (equal outcomes across groups), equalized odds (equal error rates), calibration (equal accuracy per group), or group fairness (equal treatment of similar individuals). By quantifying fairness, researchers, developers, and regulators can identify discrimination, track bias reduction efforts, and establish acceptable fairness baselines.

The primitive emerged from Gender Shades research documenting that commercial facial recognition systems had error rates of 0.8% for light-skinned males but 34% for dark-skinned females—demonstrating that apparently working systems could systematically harm marginalized communities. Fairness metrics formalize what Gender Shades revealed: apparent system accuracy masks disparate impact. By measuring fairness for each demographic group, discrimination becomes visible and measurable.

Fairness metrics face inherent tensions (different fairness definitions may be mathematically impossible to satisfy simultaneously) and require judgment about which fairness concept to prioritize. This makes fairness metrics tools for transparency and deliberation rather than technical solutions, requiring community input into which fairness definitions matter.

Technical Specifications

• Primary Function: Quantify algorithmic bias and disparate impact across demographic groups using standardized mathematical fairness measures
• Technology Stack: Statistical software for bias calculation; fairness libraries (AI Fairness 360, Fairlearn, Themis-ML) implementing multiple fairness metrics; data visualization dashboards; benchmarking systems comparing fairness across algorithms; blockchain systems documenting fairness audits
• Dependencies: Demographic data enabling bias calculation (with privacy protections); agreement on which fairness metrics matter in context; regular bias assessment and reporting; governance linking fairness metrics to system modifications or retirement; community participation in defining acceptable fairness levels

Civic Applications

• Use Cases: Criminal justice algorithms (stop-and-search, bail, sentencing); hiring algorithms (resume screening); benefit allocation systems; medical diagnosis algorithms; financial services (lending decisions); public housing allocation; ensuring equitable access to services; regulatory compliance and accountability
• Examples: Gender Shades research prompting federal investigation of facial recognition; regulatory requirements for fairness assessment in EU AI Act; ProPublica COMPAS investigation documenting racial bias in criminal sentencing algorithm; healthcare AI fairness requirements in Medicare; lending fairness regulations; algorithmic audits in Danish municipalities; insurance fairness standards