Justifiable Fairness in AI

• Justifiable Fairness is a framework in AI that requires fairness claims to be supported by normative, evidentiary, and contextual justifications beyond mere statistical metrics.
• It bridges theoretical justice concepts, such as Rawlsian principles and anti-universalism, with practical AI applications by focusing on who benefits and who is burdened.
• It operationalizes fairness through formal decision trees, causal models, and accountability mechanisms, ensuring that trade-offs and interventions are publicly contestable.

Searching arXiv for recent and foundational work on “justifiable fairness” and closely related fairness-justification frameworks. arXiv_search({"^^^^^^^^6query6 fairness\"^^^^^^^^6 OR ti:\6^^^^^^^^"Navigating Fairness Measures and Trade-Offs\"^^^^^^^^6 OR ti:\6^^^^^^^^"Is calibration a fairness requirement?\"^^^^^^^^6 OR ti:\6^^^^^^^^"Am I Being Treated Fairly?\"^^^^^^^^6 OR ti:\6^^^^^^^^"Towards the Right Kind of Fairness in AI\"","max_results":^^^^6all:\6query6^^^^,"sort_by":"relevance"}) Justifiable fairness is a family of approaches in which fairness in AI, automated decision-making, and related allocation systems is not established merely by satisfying a statistical constraint. A fairness claim is justified only when the choice of metric, the treatment of trade-offs, and the resulting allocation of benefits and burdens can be defended by substantive reasons about justice, discrimination, welfare, opportunity, evidence, or institutional legitimacy. Across the literature, the term appears in several overlapping forms: as a Rawlsian framework for selecting fairness measures, as a context-sensitive account of wrongful discrimination, as a causal criterion restricting sensitive influence to admissible pathways, and as a procedural requirement that decisions be intelligible, contestable, and verifiable to those affected (&&&^^^^6query6^^^^&&&, &&&^^^^6all:\6^^^^&&&, &&&^^^^6 OR ti:\6^^^^&&&, &&&^^^^6 OR ti:\6^^^^&&&).

^^^^6all:\6^^^^. From metric satisfaction to justified fairness claims

A central premise of the literature is that fairness metrics are indispensable but underdetermining. Statistical criteria such as demographic parity, equal opportunity, predictive equality, calibration, or equalized odds detect different kinds of disparity, yet they do not by themselves determine which notion of fairness matters in a given application. This is partly because the field contains multiple incompatible fairness definitions, often organized around independence, sufficiency, and separation, and partly because realistic settings typically do not permit simultaneous satisfaction of all attractive criteria (&&&^^^^6 OR ti:\6^^^^&&&, &&&^^^^6query6^^^^&&&).

This underdetermination is not merely technical. Several works argue that once fairness is framed as an optimization problem, the decisive question becomes normative: which departures from baseline prediction, accuracy, priority, or procedure are morally acceptable, and for whom. The Rawlsian argument in “Navigating Fairness Measures and Trade-Offs” treats fairness metrics as operational proxies whose relevance depends on how they affect the least advantaged and on whether the decision concerns access to socially important opportunities (&&&^^^^6query6^^^^&&&). The anti-universalist argument in “Is calibration a fairness requirement?” rejects the idea that any single metric, including between-group calibration, is always required; whether calibration or error-rate parity is fair depends on the moral structure of the decision context and on whether a socially salient group is made worse off relative to another (&&&^^^^6all:\6^^^^&&&). The Goodhart-oriented critique in “Are There Exceptions to Goodhart’s Law? On the Moral Justification of Fairness-Aware Machine Learning” extends the point: a fairness metric may correlate with a fair distribution of benefits and burdens under specific assumptions, yet cease to track that moral object once it becomes an optimization target (Weerts et al., 2022).

Accordingly, justifiable fairness is best understood not as a new metric but as a second-order demand on fairness practice. It asks why a given criterion should count as fair here, what assumptions make that criterion appropriate, and whether the intervention that optimizes it actually advances the moral purpose that motivated its selection. This shifts fairness from metric compliance to reason-giving.

^^^^6 OR ti:\6^^^^. Substantive normative accounts

The substantive literature links justifiable fairness to distinct theories of justice and discrimination. In Rawlsian form, fairness is justified when a metric serves either fair equality of opportunity or the difference principle. Buijsman’s framework begins from Rawls’ lexical ordering of equal basic liberties, fair equality of opportunity, and benefit to the least advantaged, and translates this into a decision rule for AI systems: if the system governs access to “offices and positions in society” or comparable opportunities, parity-based measures, especially conditional demographic parity, function as practical proxies; otherwise, the relevant metric is the one tied to the model behavior that most affects the absolute welfare of the least advantaged group (&&&^^^^6query6^^^^&&&).

The statistical criteria discussed in this literature are familiar but morally non-equivalent. Demographic parity requires

PRESERVED_PLACEHOLDER_^^^^6query6^^^^

equal opportunity requires

PRESERVED_PLACEHOLDER_^^^^6all:\6^^^^

and predictive equality requires

PRESERVED_PLACEHOLDER_^^^^6 OR ti:\6^^^^

Between-group calibration, or sufficiency, is expressed as

PRESERVED_PLACEHOLDER_^^^^6 OR ti:\6^^^^

The point of justifiable fairness is that none of these equalities is morally self-interpreting (&&&^^^^6query6^^^^&&&, &&&^^^^6all:\6^^^^&&&).

A different substantive account identifies fairness with the absence of prima facie wrongful discrimination. On this view, developed by Loi and Heitz, a violation of calibration may be unfair in some cases and not unfair in others, because the relevant question is whether the practice makes members of a socially salient group worse off relative to another through the relevant act type (&&&^^^^6all:\6^^^^&&&). The analysis is comparative and welfare-sensitive. A recommendation system can make calibration morally central because differential score meaning alters users’ self-regarding decisions, while a parole system can make equalized error burdens morally central because the harms arise through detention decisions rather than the semantic equality of scores.

The utility-and-entitlement framework of Weerts and collaborators pushes this further by locating fairness in the distribution of benefits and burdens across Potential Space, Construct Space, Observed Space, Decision Space, and Utility Space. On this view, whether demographic parity or equalized odds is justified depends not only on the causes of inequality but also on what decision subjects have a moral claim to receive or avoid, and on whether the relevant utility should track realized condition, underlying potential, or some other space (Weerts et al., 2022). Justifiable fairness thus depends on a defended account of entitlements and harm, not solely on statistical symmetry.

An adjacent but distinct contribution is “accurate fairness,” which redefines individual fairness so that an individual and her “similar counterparts” must receive predictions conforming to a uniform target, namely the ground truth of the original individual. Formally, for counterparts PRESERVED_PLACEHOLDER_^^^^6 OR ti:\6^^^^,

PRESERVED_PLACEHOLDER_^^^^6 OR ti:\6^^^^

The aim is to rule out both “true bias” and “false fairness”: predictions should not differ solely because sensitive attributes vary, but fairness should also remain anchored to truth-aligned distinctions (&&&^^^^6all:\6 OR ti:\6^^^^&&&). This suggests a truth-aligned version of justifiable fairness, though its force depends on the validity of the labels.

^^^^6 OR ti:\6^^^^. Procedural, epistemic, and contestatory dimensions

A major strand of the literature argues that fairness is not fully justified unless it can be inspected, challenged, and verified by those subject to the decision. “Am I Being Treated Fairly? A Conceptual Framework for Individuals to Ascertain Fairness” recasts fairness as an epistemic right of an individual to attain information about decisions and use that information to contest and seek effective redress against those decisions. Its four-part framework combines fairness of predictions, fairness of recourses, a contestation mechanism, and a reporting or audit mechanism. The central distinction is between explanation and justification: explanations may show how a decision was produced, but determining whether treatment is justified requires reasons, accountable entities, contestation, and possible redress (&&&^^^^6 OR ti:\6^^^^&&&).

This procedural turn is reinforced by work on evidence-based justification. “Justifiable Artificial Intelligence” argues that in high-stakes legal applications, especially with LLMs, the urgent requirement is not full internal explainability but the presentation of evidence from trustworthy sources for and against a claim. Under this view, justification shifts from model introspection to claim validation: outputs become more trustworthy, and more contestable, when they are accompanied by supporting and contradicting material that a human can inspect, validate, or reject (&&&^^^^6all:\66^^^^&&&). A related sequential-decision formulation uses a debate-based reward model, where two agents provide competing evidence for different actions and a judge proxy evaluates which action is better justified. Here justifiability becomes a learned reward signal for reinforcement learning, emphasizing concise, decision-relevant, and refutation-resistant evidence (&&&^^^^6all:\6^^^^7&&&).

The same theme appears in collective decision-making. “The Fairness Fair” argues that fair solutions should not only satisfy axioms such as envy-freeness or proportionality, but also be governed by human and societal cognition, consider perceived outcomes based on human judgment, and be verifiable. Fairness claims that remain inaccessible to non-experts, or that cannot be checked without unreasonable cognitive or informational burden, lack a crucial element of public legitimacy (&&&^^^^6all:\6^^^^8&&&). Taken together, these works enlarge justifiable fairness from a property of outputs to a property of social answerability.

^^^^6 OR ti:\6^^^^. Formal and algorithmic operationalizations

Several papers translate justifiable fairness into explicit workflows or formal mechanisms. The Rawlsian operationalization in Buijsman proceeds in six steps: determine whether the system governs access to socially important positions or opportunities; if so, use parity-based measures, especially conditional demographic parity, as operational proxies for fair equality of opportunity; if not, identify the least advantaged or most vulnerable group; determine which model outcomes or error types most affect that group’s absolute welfare; evaluate candidate metrics and fairness–accuracy trade-offs by their expected effect on that welfare; and use distributed goods or harms as provisional welfare indicators when no general welfare metric is available (&&&^^^^6query6^^^^&&&).

The “Fairness Compass” offers a parallel but more engineering-oriented selection device. It formalizes metric choice through a decision tree that asks about policy commitments, representational goals, base rates, ground truth, label bias, legitimate explanatory variables, and whether precision- or recall-sensitive harms dominate. Depending on these answers, it directs users toward independence-based, sufficiency-based, or separation-based criteria and, importantly, documents the reasoning behind the choice (&&&^^^^6 OR ti:\6^^^^&&&). In a related spirit, “Implementing Fairness in AI Classification: The Role of Explainability” argues that fairness interventions become trustworthy only when the training process, the actual consequences of the chosen criterion, and the trade-offs relative to nearby alternatives are explained. Its FairDream case study is notable because a method intended to enforce Demographic Parity empirically tended to satisfy Equalized Odds instead, forcing an explanation of why a “conservative” in-processing method stayed more closely anchored to true labels than a stricter parity-enforcing alternative (&&&^^^^6 OR ti:\6all:\6^^^^&&&).

Formalizations also appear in causal and logical work. In CausalPre, justifiable fairness is defined interventionally: a classifier $\mathcal{M}$ is justifiably fair if it is $\mathcal{K}$-fair for every superset $\mathcal{K}$ such that

$$\mathcal{A} \subseteq \mathcal{K} \subseteq \mathcal{X},$$

where PRESERVED_PLACEHOLDER_^^^^6all:\6query6^^^^ is the set of admissible attributes. The framework seeks a preprocessed distribution in which any remaining effect of sensitive attributes on the label is mediated only through admissible pathways, so that any reasonable classifier trained on the data satisfies justifiable fairness (&&&^^^^6 OR ti:\6^^^^&&&). “Toward A Logical Theory Of Fairness and Bias” reconstructs fairness through unawareness, demographic parity, and counterfactual fairness in epistemic situation calculus, making explicit what is true in the environment, what the agent knows, what observations reveal, and what hidden assumptions are needed for a fairness judgment (&&&^^^^6 OR ti:\6 OR ti:\6^^^^&&&).

Other operationalizations focus on justifications themselves. “Fairness-aware Summarization for Justified Decision-Making” extracts textual summaries PRESERVED_PLACEHOLDER_^^^^6all:\6all:\6^^^^ intended to remain sufficient for predicting PRESERVED_PLACEHOLDER_^^^^6all:\6 OR ti:\6^^^^ while minimizing demographic leakage about protected attribute PRESERVED_PLACEHOLDER_^^^^6all:\6 OR ti:\6^^^^, thereby treating fair justification as a joint condition of predictive utility and protected-attribute independence (&&&^^^^6 OR ti:\6 OR ti:\6^^^^&&&). In this line of work, justifiable fairness is implemented by reshaping the evidence basis of a decision rather than only its final label.

^^^^6 OR ti:\6^^^^. Applications and domain-specific forms

The practical force of justifiable fairness becomes visible in domain-specific applications. In lending, Buijsman’s analysis shows why a parity-improving intervention can still be unjustified if it harms the disadvantaged group in absolute terms. Simulation-based examples indicate that equalizing loan approval rates may increase unaffordable loans, defaults, or debt burdens for underprivileged borrowers; this makes false positives a morally central error type in some lending contexts, and suggests that reducing harmful false positives may be more justified than maximizing demographic parity (&&&^^^^6query6^^^^&&&). In healthcare diagnosis, by contrast, false negatives may be more harmful because missed diagnoses delay treatment, so a Rawlsian approach may prioritize equalizing false negative rates or, equivalently, true positive rates.

In multi-component recommender systems, justifiable fairness requires end-to-end analysis rather than local assurances. Wang and collaborators show that equality of exposure and pairwise ranking fairness generally do not compose: a system whose components are individually fair can still be unfair after multiplicative composition. A fairness claim becomes justifiable only if it concerns the actual composed system users experience, or if specific structural conditions guarantee that local fairness transfers to the global recommender (&&&^^^^6 OR ti:\66^^^^&&&).

In criminal justice, justifiable fairness is explicitly institutional and publicly contestable. “Alternative Fairness and Accuracy Optimization in Criminal Justice” argues that fairness cannot be abstract or universal; it must be need-based, transparent and accountable, and narrowly tailored. Its proposed technical modification replaces exact parity in false negative rates with a weighted error objective under a tolerance constraint,

PRESERVED_PLACEHOLDER_^^^^6all:\6 OR ti:\6^^^^

while making the ethical choice of error costs explicit. The paper’s larger point is that fairness claims in public systems are defensible only when the selected error weights, tolerances, and trade-offs can be explained as legitimate public choices rather than hidden technical defaults (&&&^^^^6 OR ti:\6^^^^7&&&).

A structurally different but conceptually related application appears in matching markets. “Justifiable Priority Violations” asks when departures from school priorities are fair enough to permit Pareto improvements beyond Deferred Acceptance. A priority violation is justifiable when the affected student either directly benefits from the new assignment or is unimprovable under any assignment that Pareto-dominates DA. Here justifiable fairness means that departures from entitlement are permissible only when every harmed claim can be defended to the harmed student (&&&^^^^6 OR ti:\6^^^^8&&&).

6. Enduring tensions and open problems

The literature is unified less by a single solution than by a common diagnosis: fairness claims are fragile unless their normative and epistemic premises are made explicit. Multiple tensions remain unresolved. First, the welfare consequences of fairness interventions are often difficult to estimate, especially over long horizons. Rawlsian approaches require evaluating absolute effects on the least advantaged, but the relevant welfare measure may be money, health outcomes, QALYs, treatment burdens, or something domain-specific, and none of these choices is value-neutral (&&&^^^^6query6^^^^&&&).

Second, label validity remains a recurrent fault line. Truth-aligned and error-parity approaches presuppose that labels are informative and normatively acceptable, but several papers emphasize that labels may encode measurement bias, background injustice, or historically distorted institutional practice (Weerts et al., 2022, &&&^^^^6 OR ti:\6all:\6^^^^&&&). Causal approaches alleviate some of this by distinguishing admissible from inadmissible pathways, yet they depend on the defensibility of that partition and, in practical settings, on approximations rather than complete causal knowledge (&&&^^^^6 OR ti:\6^^^^&&&).

Third, procedural justification is necessary but not sufficient. Evidence-backed outputs, contestability, auditability, and debate-based reason-giving can make systems more accountable, but they do not guarantee fair treatment if the evidence base is biased, the judge proxy is misaligned, the explanatory interface is inaccessible, or one group is disproportionately exposed to uncertainty or deferment (&&&^^^^6 OR ti:\6^^^^&&&, &&&^^^^6all:\66^^^^&&&, &&&^^^^6all:\6^^^^7&&&). The same caution applies to human-centered verification: fair solutions may need to be verifiable and acceptable, yet perceived fairness, procedural satisfaction, and formal guarantees do not always coincide (&&&^^^^6all:\6^^^^8&&&).

Finally, the literature remains pluralistic. Some accounts ground justifiable fairness in public moral reasons about the least advantaged, some in comparative disadvantage and prima facie wrongful discrimination, some in causal admissibility, some in epistemic rights and contestation, and some in domain-specific entitlement structures such as school priorities or criminal justice error costs. The cumulative lesson is not that one of these frameworks has displaced the others, but that fairness becomes justifiable only when the relevant theory of justice, discrimination, evidence, or legitimacy is made explicit enough to guide action and defend it publicly (&&&^^^^6query6^^^^&&&, &&&^^^^6all:\6^^^^&&&, &&&^^^^6 OR ti:\6^^^^&&&).