Grounding Algorithmic Fairness in Normative Baselines

40 concepts22 papers1 proposal

When machine learning practitioners build "fair" models, they typically turn to mathematical constraints: equalizing false positive rates across groups, ensuring demographic parity, or balancing predictive parity. These metrics are clean, programmable, and easily integrated into loss functions. However, a growing body of research warns that these technical fixes suffer from normative opacity—they hide profound moral and social assumptions behind a facade of mathematical neutrality.

The Illusion of Mathematical Neutrality

As argued by Hilde J. P. Weerts et al. in Algorithmic Unfairness Through the Lens of EU Non-Discrimination Law, formal equality is an inherently "empty" concept until a normative baseline is established. To claim that two individuals or groups should be treated equally, we must first make a value-based decision about which characteristics are relevant for comparison.

Without explicitly justifying this baseline, technical interventions rely on unstated social assumptions that may fail to meet legal or ethical standards. This conceptual vagueness means that metrics can become mathematically optimized yet completely disconnected from the actual social harms they are meant to mitigate, a vulnerability highlighted by Solon Barocas et al. in Language (Technology) is Power: A Critical Survey of "Bias" in NLP.

Mapping Metrics to Distributive Justice

To move past ad-hoc metrics, researchers are bridging the gap between machine learning and political philosophy. The core intuition is that every fairness metric is an implicit commitment to a specific theory of distributive justice—the branch of philosophy and economics concerned with how benefits and burdens should be allocated in a society.

Consider the economic framework of Equality of Opportunity (EOP). In economic theory, an individual's outcome is driven by two sets of factors: circumstances (systemic advantages or protected attributes like race or gender, for which they should not be held accountable) and effort (choices within their control, for which they can be held accountable).

[Algorithmic Metric] ──(Implicitly Maps To)──> [Distributive Justice Framework]
                                                        │
                                           ┌────────────┴────────────┐
                                           ▼                         ▼
                                    [Circumstances]               [Effort]
                                 (Systemic/Protected)       (Individual Control)

In A Moral Framework for Understanding Fair ML through Economic Models of Equality of Opportunity, Hoda Heidari et al. demonstrate that common algorithmic fairness metrics, such as Equalized Odds, directly map to these economic models. However, because developers rarely make these philosophical commitments explicit, they inadvertently automate assumptions about what constitutes "deservingness" or "effort." As Krishna P. Gummadi et al. note in Economic Theories of Distributive Justice for Fair Machine Learning, this lack of transparency prevents stakeholders from evaluating whether a model's mathematical behavior aligns with the actual ethical goals of the domain.

When we select a metric blindly, we also ignore alternative patterns of justice. For instance, Joachim Baumann et al. in Distributive Justice as the Foundational Premise of Fair ML: Unification, Extension, and Interpretation of Group Fairness Metrics and Unification, Extension, and Interpretation of Group Fairness Metrics for ML-Based Decision-Making point out that standard metrics fail to account for philosophies like prioritarianism or maximin (which seek to maximize the welfare of the least advantaged group). Similarly, Shrimai Prabhumoye et al. in Principled Frameworks for Evaluating Ethics in NLP Systems show that evaluating a system through a Kantian deontological lens versus a utilitarian lens yields fundamentally conflicting requirements for what makes a model ethical.

The Limits of Explainability as a Fairness Remedy

A popular intuition in AI research is that Explainable AI (XAI) can serve as a natural bridge to fairness. The logic seems straightforward: if an explanation tool (like feature attribution or counterfactuals) reveals why a model made a decision, developers can identify and eliminate bias.

However, the authors in this space show that this connection is largely unproven and overextended. In A Critical Survey on Fairness Benefits of Explainable AI, Luca Deck et al. reveal that most claims linking XAI to fairness are vague and treat fairness as a monolithic concept. They rarely specify the exact mechanism of contribution or the intended beneficiaries, ignoring the technical limitations of post-hoc explanations.

This theoretical gap is echoed across several studies:

Goodhart's Law and the Shift to Human-Centric Grounding

What happens when we force a model to optimize for a normatively ungrounded mathematical metric? We run directly into Goodhart's Law: when a measure becomes a target, it ceases to be a good measure.

In Are There Exceptions to Goodhart's law? On the Moral Justification of Fairness-Aware Machine Learning, Hilde Weerts et al. demonstrate that treating fairness metrics as hard optimization constraints can paradoxically result in net harm or deeply unfair outcomes for individuals because the math ignores the real-world utility of the decisions.

To resolve this, the field is shifting toward descriptive ethics and procedural fairness—examining how humans actually reason about fairness rather than prescribing abstract formulas. For example, Krishna P. Gummadi et al. in Human Perceptions of Fairness in Algorithmic Decision Making: A Case Study of Criminal Risk Prediction discover that human fairness judgments are driven by latent properties of features—such as relevance, privacy, and volitionality—rather than statistical parity.

When technical workflows ignore these human elements, they create new ethical dilemmas. For instance, Raymond Feng et al. in Adapting Fairness Interventions to Missing Values show that while treating missing data patterns as a predictive signal can improve group-level metrics, it introduces serious concerns regarding procedural fairness and the justifiability of differential treatment. Similarly, using generative AI to balance datasets, as evaluated by Benoît Ronval et al. in Can Generative Ai-based Data Balancing Mitigate Unfairness Issues in Machine Learning?, serves as a descriptive tool but lacks the deep philosophical justification required to prove absolute fairness.

Ultimately, as shown by Alessandro Fabris et al. in Dissatisfaction Induced by Pairwise Swaps, fairness metrics only become meaningful when they are explicitly grounded in behavioral realities—such as reframing abstract ranking formulas into measurable human experiences like content producer dissatisfaction.

Key Foundational Work

A Moral Framework for Understanding Fair ML through Economic Models of Equality of Opportunity (2019) — contributed 1 concept elements to this direction.
Principled Frameworks for Evaluating Ethics in NLP Systems (2019) — contributed 1 concept elements to this direction.
Human Perceptions of Fairness in Algorithmic Decision Making: A Case Study of Criminal Risk Prediction (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_normative_baseline_dependency_4ac8d2b2,
  title = {normative baseline dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/normative-baseline-dependency/rdr_4ac8d2b2ed514ca281d9aed70fa85d51},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_normative_baseline_dependency_4ac8d2b2_v1,
  title = {normative baseline dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/normative-baseline-dependency/rdr_4ac8d2b2ed514ca281d9aed70fa85d51/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake