Distributive Justice

An egalitarian approach to distributive justice demands that an AI system distribute benefits and burdens equally across all individuals, regardless of other characteristics. In a personalization engine, this translates to enforcing strict demographic parity where every user segment receives identical offer rates or exposure.

• Mechanism: Post-processing thresholds or constrained optimization
• Trade-off: Often reduces overall business utility to achieve perfect equality
• Example: A coupon engine that ensures every ZIP code receives the same discount rate, ignoring local demand elasticity

Prioritarianism weights benefits more heavily when they accrue to those who are worse off in a relevant dimension. In AI systems, this means a recommendation engine might deliberately boost exposure for disadvantaged sellers or offer steeper discounts to low-engagement users.

• Implementation: Weighted loss functions that penalize under-serving disadvantaged groups
• Key Metric: Maximizing minimum welfare (maximin principle)
• Use Case: A marketplace algorithm that prioritizes listings from small, minority-owned businesses in search rankings

A utilitarian framework seeks to maximize total aggregate welfare across the entire population, even if that produces unequal individual outcomes. This is the default posture of most unconstrained recommender systems that optimize for global click-through rate or revenue.

• Mechanism: Standard loss minimization without fairness constraints
• Risk: Can systematically ignore minority groups if they represent low-revenue segments
• Critique: Just because total utility increases does not mean the distribution is just—a system could double revenue for 90% of users while completely excluding 10%

Desert theory ties the fair distribution of outcomes to an individual's past actions, effort, or merit. In personalization, this manifests as loyalty-based tier systems where long-tenured or high-spend customers receive preferential treatment.

• Implementation: Feature engineering that encodes historical contribution as a primary input
• Fairness Challenge: Historical merit may correlate with systemic advantage, perpetuating inequality
• Example: An airline upgrade algorithm that prioritizes frequent flyers—fair by contribution, but potentially regressive

Derived from John Rawls's theory of justice, the difference principle permits inequalities only if they benefit the least advantaged members of society. An AI system following this rule would allow a high-spend segment to receive better offers only if that surplus demonstrably improves outcomes for the lowest-engagement cohort.

• Constraint: Inequalities must be justified by reciprocal benefit to the worst-off group
• Audit Mechanism: Requires causal analysis linking preferential treatment of one group to measurable gains for another
• Application: A dynamic pricing model that charges premium users more, using the margin to subsidize access for low-income users

Sufficientarianism focuses on ensuring every individual reaches a minimum adequate threshold of benefit, rather than enforcing strict equality. Once all users meet this baseline, inequalities above the threshold are considered ethically permissible.

• Implementation: Constrained optimization with a hard floor on per-group performance metrics
• Key Distinction: Differs from egalitarianism by allowing variance above the sufficiency line
• Example: A credit scoring model that guarantees a minimum approval rate for every demographic, but allows variation beyond that floor based on risk

A group fairness metric requiring a model's positive prediction rate to be equal across all demographic groups. This enforces statistical independence from the sensitive attribute, ensuring that the proportion of users receiving a favorable outcome is identical regardless of group membership. In practice, this is often the first metric audited in high-stakes allocation scenarios like loan approvals or promotional offer distribution.

A stricter fairness criterion requiring a classifier to have equal true positive rates and false positive rates across protected groups. Unlike demographic parity, this metric accounts for ground-truth outcomes, ensuring that errors are evenly distributed. It prevents a system from being accurate for one group while making systematic mistakes for another, which is critical in risk assessment and medical resource allocation.

The inherent tension where enforcing strict fairness constraints often results in a measurable reduction in predictive accuracy or business utility. Distributive justice in AI requires navigating this Pareto frontier explicitly. Key considerations include:

• Quantifying the cost of fairness in terms of revenue or precision
• Selecting a socially acceptable operating point on the trade-off curve
• Using multi-objective optimization to surface non-dominated solutions

A causal definition of fairness where a decision is considered fair if it would remain the same in a counterfactual world where an individual's sensitive attributes were different. This approach uses structural causal models to distinguish legitimate causal pathways from discriminatory ones. It is particularly relevant for distributive justice as it ensures outcomes are based on meritocratic factors rather than protected characteristics.

A fairness criterion ensuring that a model's predicted probabilities accurately reflect the true likelihood of an outcome for every distinct demographic group. A well-calibrated system means that when it predicts a 70% probability of a positive outcome, that outcome materializes 70% of the time for each group. This prevents systematic over- or under-estimation of risk that could lead to inequitable resource distribution.

A framework for multi-stakeholder platforms that simultaneously optimizes for equitable outcomes on both sides of a marketplace. In retail personalization, this means balancing fair exposure for product suppliers with fair recommendations for consumers. This concept extends distributive justice beyond individual users to consider the entire ecosystem of affected parties, preventing the optimization of one group's outcomes at the expense of another's.