# Algorithmic Fairness

Algorithmic fairness evaluates and addresses unjust differences in how automated systems allocate benefits, impose burdens, or represent people.

- Type: principle
- Domain: Data, experimentation & growth
- Level: Advanced
- Last reviewed: 2026-09-26
- Original: [Soroush’s Product Wiki](<https://wiki.soroush.uk/topic/algorithmic-fairness/>)

## What it means

Algorithmic fairness concerns the treatment and outcomes of people affected by algorithmic systems. Assessment may compare selection rates, error rates, or prediction quality across relevant groups, but a fairness metric encodes a particular concern rather than a universal definition of justice. Harm can arise from the task, labels, data, model, interface, or surrounding institution, even without discriminatory intent.

## Why it matters

Good average performance can conceal repeated exclusion or unequal errors. Product teams need to identify who benefits, who bears mistakes, and which disparities matter before choosing an optimization objective or release threshold.

## When to use it

Assess fairness when automated ranking, allocation, classification, or representation affects people's opportunities or experiences. Involve affected groups and appropriate specialists when defining harms and handling sensitive attributes. Use qualitative investigation alongside quantitative measures when impacts cannot be captured by available labels.

## How it works

1. Define the decision, affected groups, potential harms, and relevant context with stakeholders.
2. Examine how labels, sampling, missing data, proxies, and workflow choices may create or conceal disparities.
3. Choose measures tied to the harm, such as selection-rate differences or false-positive and false-negative rates, with uncertainty estimates.
4. Compare groups and meaningful intersections while checking small samples and the validity of the reference labels.
5. Evaluate changes to data, model, thresholds, or workflow, document tradeoffs, and monitor outcomes and recourse.

## A practical example

A community platform finds that a moderation classifier incorrectly removes posts in one dialect more often than others. Reviewers inspect the labels and context, the team expands evaluation coverage, and it tests changes to the model and appeal process rather than optimizing only aggregate accuracy.

## Where it breaks

Different fairness criteria can conflict, and equal measured rates do not prove the absence of harm. Group definitions and historical labels can be contested or incomplete. Sensitive data requires careful governance, and small samples can make apparent differences unstable.

## Common mistakes

- Declaring a system fair because protected attributes were removed while proxies remain.
- Selecting whichever fairness metric makes current results look best.
- Treating historical human decisions as unbiased ground truth.
- Ignoring intersectional groups, uncertainty, or harms outside the measured decision.

## Visualization

[View the visualization on the website](<https://wiki.soroush.uk/topic/algorithmic-fairness/#visualization-algorithmic-fairness>)

## Sources

1. [Towards a Standard for Identifying and Managing Bias in Artificial Intelligence](<https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.1270.pdf>) — National Institute of Standards and Technology · Government research report
2. [Common fairness metrics](<https://fairlearn.org/main/user_guide/assessment/common_fairness_metrics.html>) — Fairlearn · Official documentation

## Related topics

- [Model Evaluation](<https://wiki.soroush.uk/topic/model-evaluation/>)
- [Responsible Product Management](<https://wiki.soroush.uk/topic/responsible-product-management/>)
- [Dataset Card](<https://wiki.soroush.uk/topic/dataset-card/>)
