Algorithmic Fairness through the Lens of Social Welfare

This direction is split into other directions: consequentialist algorithmic fairness. Its last version stays readable and citable here.

Version 1 of 117 papers28 conceptsfrozen 2026-10-04

In the pursuit of algorithmic fairness, researchers have traditionally relied on mathematical parity: ensuring that error rates or selection frequencies are equal across demographic groups. However, a growing body of research suggests that these statistical constraints can be a blunt instrument. If a model achieves "fairness" by reducing the quality of service for everyone—a phenomenon known as "leveling down"—it may satisfy a mathematical equation while leaving the very groups it intended to protect worse off.

To solve this, a new direction in machine learning shifts the focus from statistical parity to social welfare. By treating algorithmic predictions as a problem of resource allocation and human utility, this framework ensures that fairness interventions actually improve the well-being of marginalized groups.

The Failure of Mathematical Parity

The core tension in traditional fairness is the "fairness-utility trade-off." As Lily Hu and Yiling Chen argue in Fair Classification and Social Welfare, tightening fairness constraints can sometimes violate the Pareto Principle, a fundamental economic concept where a change is only considered an improvement if it makes at least one person better off without making anyone else worse off. In some cases, enforcing strict statistical parity forces a "Planner" to reject qualified candidates from all groups just to keep the ratios even.

This happens because standard metrics treat all errors as equal. But in the real world, a false negative (denying a loan to someone who could repay it) has a different human cost than a false positive. Hoda Heidari and Krishna P. Gummadi point out in A Moral Framework for Understanding Fair ML through Economic Models of Equality of Opportunity that we must define utility functions—mathematical representations of the benefit or harm an individual derives from an outcome—to understand the true moral weight of an algorithm's decision.

From Parity to Welfare-Equalization

If the goal is to help a disadvantaged group, the objective should not be to match the error rates of the advantaged group, but to equalize the expected utility.

Omer Ben-Porat, Fedor Sandomirskiy, and Moshe Tennenholtz formalize this in Protecting the Protected Group: Circumventing Harmful Fairness. They introduce the Welfare-Equalizing (WE) framework, which proves that as long as the regulator and the population agree on which group is disadvantaged, focusing on utility ensures that fairness interventions are "helpful" rather than "harmful."

This shift changes the optimization problem:

  1. Traditional Fairness: Minimize error subject to $P(\hat{Y}=1 | G=A) = P(\hat{Y}=1 | G=B)$.
  2. Welfare Fairness: Maximize the utility (or "happiness") of the groups, ensuring the disadvantaged group receives a fair share of the realized value.

In Happiness as a Measure of Fairness, Georg Pichler and Pablo Piantanida demonstrate that this "happiness" can be calculated using linear programming, making it a computationally efficient way to post-process models for equitable outcomes.

Distributive Justice and the Helpfulness Gap

This research direction draws heavily from political philosophy, specifically Distributive Justice. Joachim Baumann, Corinna Hertweck, and Michele Loi argue in Distributive Justice as the Foundational Premise of Fair ML that algorithmic fairness is essentially a question of how well-being is distributed.

This perspective is particularly useful for modern AI systems like LLMs, where "fairness" isn't just about avoiding toxicity, but about the equitable distribution of benefits. For example, Jiao Sun and colleagues found a "helpfulness gap" in Are Akpans Trick or Treat: Unveiling Helpful Biases in Assistant Systems, where AI assistants provided significantly more useful information for queries related to developed nations than for less-developed ones. By framing this as a utility disparity, researchers can audit models for systemic exclusion that traditional error-rate metrics would miss.

Solving the Fairness Paradox

A welfare-based approach also helps resolve the Fairness Paradox. In Searching for Unfairness in Algorithms' Outputs, Ian Davidson and S. S. Ravi show that algorithms designed to be fair on individual attributes (like race or gender) often inadvertently increase unfairness for intersectional groups (like "Hispanic Women").

By adopting inequality measures from economics—such as the Gini Index used by Enmao Diao in Toward Unifying Group Fairness Evaluation from a Sparsity Perspective—researchers can treat model outcomes as a finite resource. This allows for a unified evaluation that satisfies the Pigou-Dalton principle: a transfer of utility from a high-performing group to a low-performing group should always improve the overall fairness score.

Ultimately, this direction moves the field toward a more human-centric goal. As Corinna Hertweck and Christoph Heitz suggest in A Justice-Based Framework for the Analysis of Algorithmic Fairness-Utility Trade-Offs, the task of the AI researcher is not just to balance equations, but to explicitly define what constitutes "well-being" and ensure that the algorithm's distribution of that well-being aligns with our collective moral values.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_algorithmic_fairness_through_the_lens_of_f093cd2c,
  title = {Algorithmic Fairness through the Lens of Social Welfare},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/algorithmic-fairness-through-the-lens-of-social-welfare/rdr_f093cd2c1d154e8b99a3f83a9f870a17},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_algorithmic_fairness_through_the_lens_of_f093cd2c_v1,
  title = {Algorithmic Fairness through the Lens of Social Welfare},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/algorithmic-fairness-through-the-lens-of-social-welfare/rdr_f093cd2c1d154e8b99a3f83a9f870a17/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}