Distributive Justice

The principle that all individuals should receive equal shares of a resource or outcome, regardless of their contribution, need, or merit. In algorithmic contexts, egalitarianism translates to strict demographic parity—requiring that a model's positive prediction rate be identical across all protected groups.

• Strict Equality: Every group receives the exact same allocation
• Algorithmic Application: Enforces equal acceptance rates in hiring or lending models
• Critique: Can ignore legitimate differentiating factors like qualification or risk
• Example: A university admissions algorithm that admits exactly 50% of applicants from each demographic group

A welfare-based theory asserting that benefiting the worst-off members of society should receive greater moral weight than benefiting those already advantaged. In machine learning, this manifests as Rawlsian max-min fairness—optimizing model performance specifically for the most disadvantaged subgroup.

• Moral Weighting: Gains for the disadvantaged count more than equivalent gains for the advantaged
• Algorithmic Application: Minimizing the maximum error rate across groups rather than optimizing average accuracy
• Distinction from Egalitarianism: Does not demand equality, only priority for the worst-off
• Example: A healthcare triage algorithm that prioritizes diagnostic accuracy for patients with historically poor health outcomes

The view that justice requires ensuring every individual reaches a minimum threshold of well-being or resource access, rather than pursuing perfect equality. Once the threshold is met, inequalities above it are morally unproblematic.

• Threshold Focus: The goal is adequacy, not equality
• Algorithmic Application: Guaranteeing a minimum acceptable model performance floor for every demographic group
• Practical Appeal: Aligns with regulatory concepts like the four-fifths rule in disparate impact law
• Example: A credit scoring model that ensures at least 70% approval rate for every protected group, even if some groups receive higher rates

A meritocratic principle holding that resources should be distributed according to what individuals deserve based on their actions, efforts, or contributions. In algorithmic fairness, this aligns with calibration—ensuring that a model's predicted probabilities accurately reflect actual outcomes for each group.

• Merit Orientation: Rewards proportional to relevant qualifications or performance
• Algorithmic Application: Ensuring equal positive predictive value across groups—if a model predicts success, the probability of actual success should be identical regardless of group membership
• Tension with Equality: Can perpetuate historical inequalities if 'merit' reflects past discrimination
• Example: A hiring algorithm that ranks candidates by job-relevant skills, where a score of 80/100 means the same likelihood of job success for all demographic groups

A framework emphasizing individual liberty and voluntary exchange, holding that distributions are just if they arise from free transactions without force or fraud, regardless of resulting inequality. In AI governance, this manifests as procedural fairness—focusing on the legitimacy of the decision-making process rather than outcome patterns.

• Process Over Outcome: Justice resides in the rules of the game, not the final score
• Algorithmic Application: Ensuring informed consent, transparency, and the right to opt-out of automated decisions
• Regulatory Alignment: Supports right to explanation provisions in GDPR and the EU AI Act
• Example: A dynamic pricing algorithm that charges different prices but allows all users to see the pricing logic and choose alternatives freely

The consequentialist principle that resources should be allocated to maximize total welfare or utility across the entire population, even if this produces significant inequalities. In machine learning, this corresponds to optimizing for overall accuracy without group-specific constraints.

• Aggregate Maximization: The greatest good for the greatest number
• Algorithmic Application: Training models to minimize total prediction error, accepting that some groups may experience higher error rates
• Conflict with Fairness: Can systematically disadvantage minority groups if they represent a small portion of the training data
• Example: A loan default prediction model optimized for overall profit that achieves 95% accuracy overall but only 70% accuracy for a small demographic subgroup

Algorithmic systems increasingly mediate access to scarce public goods such as social housing, school admissions, and healthcare triage. Distributive justice frameworks evaluate whether these systems allocate resources according to morally defensible criteria—such as need, merit, or equality—rather than optimizing solely for efficiency. For example, a kidney transplant allocation algorithm must balance medical urgency, tissue compatibility, and waiting time, embodying a hybrid distributive principle that combines prioritarianism (prioritizing the worst-off) with utilitarian considerations.

Matching algorithms in school choice, refugee resettlement, and organ donation must satisfy competing distributive claims. The Gale-Shapley deferred acceptance algorithm guarantees stable matches but raises distributive questions: whose preferences are prioritized? A student-optimal versus school-optimal matching embodies different distributive commitments. More broadly, envy-freeness—the property that no agent prefers another's allocation to their own—serves as a formal distributive justice criterion in algorithmic mechanism design.

Predictive algorithms in criminal justice, credit scoring, and insurance distribute risks and burdens across populations. A recidivism prediction model that generates higher false positive rates for certain groups imposes a disproportionate risk burden—the cost of being wrongly classified. Distributive justice demands scrutiny of who bears these error costs. The equalized odds fairness criterion directly addresses this by requiring equal error rates across groups, operationalizing a distributive principle that error burdens should not correlate with protected characteristics.

Distributive justice extends beyond model outputs to the distribution of AI's economic gains. When foundation models are trained on publicly available data, questions arise about who should benefit from the resulting systems. Data dividends, universal basic services funded by AI productivity gains, and community-based data trusts represent institutional mechanisms for distributing the benefits of AI more broadly. This application shifts the distributive frame from individual allocation decisions to the structural distribution of technological surplus.

During crises, algorithmic triage systems must operationalize scarcity ethics under extreme constraints. Pandemic ventilator allocation protocols, disaster response routing, and emergency service dispatch all encode distributive principles. The sequential organ failure assessment (SOFA) score, used in ICU triage, exemplifies a needs-based distributive logic. Contrasting approaches—first-come-first-served, lottery, or maximizing life-years saved—reflect different underlying theories of justice, from libertarian to utilitarian to egalitarian frameworks.

Climate models and sustainability algorithms make allocation decisions with temporal scope, distributing costs and benefits across generations. Carbon pricing algorithms, resource extraction optimization, and infrastructure planning systems embed discount rates that implicitly weigh present welfare against future welfare. A high discount rate privileges current consumption, while a near-zero rate reflects a stronger commitment to intergenerational equity. This application connects algorithmic design to Rawlsian just savings principles and the ethics of long-term decision-making under uncertainty.

The systematic study and engineering practice of designing machine learning systems that make decisions without unjustified discrimination against individuals or groups based on protected attributes. It operationalizes distributive justice by translating ethical principles into mathematical constraints and quantitative metrics that can be enforced during model training and evaluation. Key approaches include pre-processing, in-processing, and post-processing interventions.

A fairness metric requiring that the probability of a positive prediction is equal across all demographic groups, regardless of true underlying outcome rates. Also known as demographic parity, this criterion directly addresses distributive justice concerns by ensuring that algorithmic benefits—such as loan approvals or job interview selections—are distributed proportionally across protected groups. Critics note it may ignore legitimate qualification differences.

A fairness criterion requiring a classifier to achieve equal true positive rates and equal false positive rates across different protected groups. Unlike statistical parity, equalized odds conditions on the true outcome, ensuring that qualified individuals have the same chance of receiving a benefit regardless of group membership. This aligns with distributive justice principles that reward merit-based allocation while penalizing systemic error disparities.

A legal doctrine and statistical measure identifying facially neutral policies or algorithms that disproportionately harm members of a protected class. Quantified through the Four-Fifths Rule: a selection rate for any group less than 80% of the highest group's rate constitutes evidence of adverse impact. This framework directly interrogates whether an allocation mechanism produces distributively unjust outcomes, even without explicit discriminatory intent.

A causal definition of fairness where a decision is considered fair if it would remain the same in a counterfactual world where the individual belonged to a different demographic group. This approach uses structural causal models to distinguish between discriminatory path-specific effects and legitimate, non-discriminatory influences. It directly addresses distributive justice by asking whether an individual's protected attribute was a cause of their outcome.

A framework for evaluating algorithmic bias that examines how overlapping social identities—such as race and gender—combine to create unique, compounded experiences of discrimination. Traditional single-axis fairness metrics often fail to detect harms at intersectional subgroups (e.g., Black women). This approach ensures distributive justice analyses capture the full complexity of how algorithmic burdens concentrate on multiply-marginalized populations.