What is Algorithmic Fairness?
Algorithmic Fairness is the sociotechnical practice of defining, measuring, and governing how an automated decision system distributes benefits, errors, and burdens across affected people under an explicit use, population, and accountability process.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Start from harms and decision responsibilities
Map who receives a benefit, burden, error, delay, or loss of autonomy, and identify who can change the data, model, policy, interface, and appeal path. NIST SP 1270 treats harmful bias as a lifecycle risk with systemic, statistical, computational, and human contributors. The useful question is not only whether scores differ, but which mechanism creates a material harm and which owner can mitigate it.
Translate the concern into an auditable estimand
Specify the decision point, eligible population, prediction horizon, favorable outcome, reference label, group definition, comparison measure, uncertainty interval, and minimum sample support. Then report task performance and data quality alongside fairness measures. The Fairlearn assessment workflow separates observed disparity from the contextual judgment about whether it is harmful or unacceptable.
Manage tradeoffs throughout the lifecycle
Different criteria can conflict, especially when outcome base rates differ and prediction is imperfect. Kleinberg, Mullainathan, and Raghavan formalize an incompatibility among calibration-style and error-balance conditions under non-degenerate settings. Treat metric selection as a documented policy decision, test thresholds and interventions before deployment, and monitor drift, overrides, complaints, appeals, and realized outcomes after release.
Key Characteristics
- Frames fairness around affected people, decisions, harms, and accountable owners
- Distinguishes data, label, model, threshold, interface, and institutional sources of disparity
- Uses metrics as context-dependent evidence rather than universal definitions of justice
- Requires explicit populations, outcomes, group definitions, time windows, and uncertainty
- Recognizes that statistical criteria can be mathematically incompatible
- Extends beyond pre-release testing to monitoring, appeals, and corrective action
Common Use Cases
- Designing a hiring-screening audit around qualified-candidate exclusion
- Evaluating lending thresholds across protected and intersectional groups
- Reviewing healthcare triage errors when labels reflect unequal access
- Comparing human overrides, complaints, and realized outcomes after deployment
- Documenting why a fairness criterion fits a particular product decision
Example
Loading code...Frequently Asked Questions
What does Algorithmic Fairness mean in practice?
It means defining which people and decisions are affected, identifying plausible harms, selecting evidence that matches those harms, testing the complete workflow, and assigning owners for mitigation and appeal. Equal group rates can be one signal, but the process also includes data quality, accessibility, human behavior, and downstream effects.
Is there one best fairness metric?
No. Demographic Parity evaluates selection without conditioning on labels, Equalized Odds compares error rates given labels, Predictive Parity compares correctness among positive predictions, and Counterfactual Fairness asks a causal individual-level question. The appropriate evidence depends on the decision, label validity, error costs, and policy goals.
Does removing protected attributes make a model fair?
Not necessarily. Other features can proxy the attribute, historical labels can encode unequal treatment, and the policy can impose unequal costs even when the model never reads the field. Protected attributes may also be needed under appropriate governance to measure disparities that would otherwise remain invisible.
Why can fairness criteria conflict?
When groups have different observed outcome prevalences and a classifier is imperfect, equalizing error rates can make positive predictive values differ, while equalizing predictive values can leave error rates unequal. The conflict is a property of the statistical conditions, not a bug that one optimizer can always remove.
Does passing a fairness audit prove legal compliance?
No. A technical audit provides evidence under a declared dataset and protocol. Legal duties depend on jurisdiction, role, use, affected rights, procedure, documentation, and current law. Teams need accountable legal and domain review, plus monitoring and contestability after deployment.