Substantive Intersectional Algorithmic Fairness

• Substantive intersectional algorithmic fairness is a framework that evaluates algorithms by considering interlocking systems of oppression and structural inequities rather than solely technical metrics.
• It emphasizes contextual specificity, relationality, and the perspectives of multiply marginalized groups through tools like subgroup audits and differential fairness measures.
• The approach challenges traditional fairness metrics by focusing on reducing upstream disparities and downstream harms while advocating for systemic reforms.

Substantive intersectional algorithmic fairness denotes an approach to algorithmic fairness that treats fairness not as a technical property of an isolated decision rule, but as a question of whether an algorithmic system reduces or reinforces interlocking structures of oppression and privilege across overlapping identities. It extends Ben Green’s account of substantive algorithmic fairness—formulated against “formal algorithmic fairness”—with intersectional feminist emphases on relationality, contextual specificity, multiply marginalized standpoints, and the social construction of categories. In this view, parity metrics, subgroup audits, and optimization constraints are auxiliary instruments rather than the definition of fairness itself; the central issue is whether algorithmic intervention reduces upstream disparities, downstream harms, and compounded disadvantage in practice (Green, 2021, Mirsch et al., 25 Aug 2025).

1. Conceptual foundations

Green characterizes formal algorithmic fairness by analogy to formal equality in law and philosophy. Formal equality requires treating like cases alike at the decision point, usually by applying the same standard and considering only characteristics deemed relevant at that moment. In algorithmic settings, this becomes a methodology focused on the input-output behavior of a particular model or decision rule, with fairness treated as a technical attribute of the system. By contrast, substantive algorithmic fairness is modeled on substantive equality: it asks whether institutional structures are producing domination, marginalization, or oppression, and whether policy should alter those structures. Green is explicit that this is not a new mathematical fairness criterion; the point is precisely to avoid folding substantive equality back into the same narrow methodology (Green, 2021).

The legal and philosophical background is central. Green grounds the substantive turn in Elizabeth Anderson’s democratic equality and relational egalitarianism, Martha Minow’s “dilemma of difference,” Joseph Fishkin’s opportunity pluralism, and Catharine MacKinnon’s work on substantive equality and social hierarchy. He also draws a legal parallel between formal equality and disparate treatment doctrine, and between substantive equality and disparate impact doctrine. This framing matters because it recasts algorithmic fairness as a problem of institutional design and social reform rather than metric selection alone (Green, 2021).

Intersectional extensions add a further claim: social hierarchies arise along lines including gender, class, ability, sexual orientation, and age, and these hierarchies often intersect and interlock. The feminist account of substantive intersectional algorithmic fairness argues that fairness cannot be reduced to adding more demographic variables to a formal audit. Instead, it requires attention to interlocking systems of oppression and privilege, situated knowledge, power, domination, and the standpoint of multiply marginalized groups. On this account, intersectionality functions both as an analytical tool and as a critical praxis (Mirsch et al., 25 Aug 2025).

2. Why formal fairness is insufficient for intersectional justice

A major target of critique is the idea that fairness can be settled at the decision point. Green’s running example is the COMPAS pretrial risk assessment controversy, where two fairness notions are in tension: separation, requiring equal false positive and false negative rates across groups, and sufficiency, requiring equal observed outcome rates within each score or prediction category across groups. Except in rare cases of perfect prediction or equal base rates, a model cannot satisfy both simultaneously. Green treats this “impossibility of fairness” not as a reason to abandon fairness, but as evidence that the field has overfocused on local decision rules while ignoring broader social context (Green, 2021).

The deeper critique is social rather than merely mathematical. Standard fairness metrics are said to fail because they treat group disparities as neutral facts rather than products of oppression, focus on the moment of decision instead of the social production of risk and disadvantage, ignore downstream harms such as detention or stigmatization, and can make oppressive institutions look legitimate if the algorithm is accurate or calibrated. Thus even a perfectly accurate model can reproduce injustice when the underlying data reflect unequal social conditions. Green’s pretrial example is explicit: if Black defendants face more criminogenic conditions because of historical and present-day oppression, then a model that accurately predicts higher risk for Black defendants may still intensify racial hierarchy by detaining more Black people (Green, 2021).

Intersectional work generalizes this point. The survey literature emphasizes that a model can appear fair on race alone and on gender alone, yet still discriminate against a combined subgroup such as Black women. This is the familiar problem of fairness gerrymandering: coarse fairness constraints can hide severe disparities in structured subgroups. A formal result in differential fairness matches that intuition: if a mechanism is fair with respect to the full protected attribute tuple $A = S_1 \times \cdots \times S_p$, then it is also fair with respect to any nonempty proper subset of those attributes; the converse does not hold (Gohar et al., 2023, Foulds et al., 2018).

3. Formal measurement repertoires within an intersectional frame

Intersectional fairness research has developed several formal repertoires for representing subgroup disparities. These tools are widely used for auditing and mitigation, but substantive accounts insist that they remain instruments for evaluating systems, not substitutes for evaluating justice itself (Green, 2021).

Framework	Core formal object	Stated role
Differential Fairness (Foulds et al., 2018)	Ratio bounds over the full Cartesian product of protected attributes	Protects full intersections and implies fairness on subsets
Worst-case comparisons (Ghosh et al., 2021)	Min/max ratios or worst pairwise subgroup ratios	Surfaces the most disadvantaged intersectional subgroup
Hypercube / levelwise variance (Filippi et al., 2023)	$3^M$-group lattice with upward propagation and $VarRatio(K)$	Diagnoses hidden bias across intersectional levels
$\alpha$-Intersectional Fairness (Maheshwari et al., 2023)	Convex tradeoff between absolute worst-group performance and relative parity	Designed to expose “leveling down”

Differential fairness is one of the most directly intersectional formal definitions. For discrete protected attributes $S_1,\ldots,S_p$ and $A = S_1 \times \cdots \times S_p$, a mechanism $M$ is $\epsilon$-differentially fair if, for all outcomes $y$ and all protected tuples $\mathbf{s}_i,\mathbf{s}_j \in A$,

$3^M$0

Smaller $3^M$1 means closer parity across all intersections, and the 80% rule corresponds to $3^M$2 (Foulds et al., 2018).

Other work adopts a Rawlsian or worst-off orientation. The worst-case comparison framework extends demographic parity, disparate impact, conditional statistical parity, equal opportunity, and related criteria by constructing Cartesian-product subgroups and then reporting either minimum-to-maximum ratios or the worst subgroup pair. In this family, fairness is assessed by the disparity experienced by the worst-off subgroup relative to the best-off subgroup (Ghosh et al., 2021).

A different formalization treats intersectionality as a geometric hierarchy on a hypercube. The “fractal” approach indexes all subgroup specifications in $3^M$3, proves that fairness at the finest intersectional level implies fairness at all coarser levels, and derives a variance benchmark under Intersectional Statistical Parity. The resulting metric family,

$3^M$4

uses levelwise variance relative to a theoretical baseline to diagnose hidden intersectional bias (Filippi et al., 2023).

A distinct critique concerns leveling down. The $3^M$5-Intersectional Fairness proposal argues that purely relative measures such as Differential Fairness can judge a model “fairer” even when the apparent gain comes from degrading the best-performing group rather than improving the worst one. It therefore combines an absolute term tracking worst-group performance and a relative term tracking inter-group parity, with $3^M$6 controlling the tradeoff (Maheshwari et al., 2023).

4. Statistical sparsity, uncertainty, and inferential caution

Intersectional auditing is not only a normative problem; it is also a severe statistical problem. As protected attributes are crossed, subgroup counts become very small. The Bayesian modeling literature describes fairness estimation in intersectional settings as a small-$3^M$7 problem: empirical subgroup rates become noisy or undefined, and zero-count pathologies can make fairness quantities infinite or unstable. The proposed remedy is to model $3^M$8 probabilistically and compute fairness metrics from posterior predictive distributions rather than raw counts. In this setting, hierarchical logistic regression provides partial pooling across related groups, and Bayesian model averaging accounts for model uncertainty; the reported result is that posterior predictive estimates are more reliable and more data-efficient than naive empirical estimates (Foulds et al., 2018).

Recent testing frameworks push the same point in a decision-theoretic direction. Size-Adaptive Fairness Testing (SAFT) replaces the usual practice of comparing a single point estimate to a fixed threshold with a size-adaptive hypothesis test. For sufficiently large subgroups it uses a Central-Limit result for subgroup statistical parity difference and a Wald test with analytic confidence intervals; for the long tail of small intersectional groups it uses a Bayesian Dirichlet-multinomial model with Monte-Carlo credible intervals. The procedure switches according to subgroup support, using $3^M$9 in the paper’s experiments, and explicitly reports intervals and a reject/fail-to-reject decision rather than a raw disparity alone (Ferrara et al., 12 Jun 2025).

A closely related moral-methodological argument is that fairness procedures should respect uncertainty without lowering standards for small groups. The “intersectionality problem” paper argues that sparse data make many point estimates statistically meaningless and, in some cases, morally misleading. Its desiderata—Minimal Justice, Consistent Conceptualization, and Incentive Compatibility—are directed against methods that relax fairness requirements for small groups or make improved data collection make a model look less fair. The paper sketches sufficiency-based hypothesis tests with a common threshold $VarRatio(K)$0, precisely to avoid group-size weighting that can become almost vacuous for very small groups (Himmelreich et al., 2024).

Further theoretical work shows why marginal auditing is not enough even statistically. Marginal fairness on each protected attribute does not in general imply intersectional fairness. Exact decomposition is available only under strong assumptions, including mutual independence of protected attributes and their conditional independence given predictions. In the general case, information-theoretic dependence measures such as total correlation and conditional total correlation yield high-probability bounds, while partition-based heuristics improve approximation under data sparsity (Molina et al., 2022).

5. Intervention logics and application domains

Substantive intersectional algorithmic fairness changes what counts as an intervention. Green’s substantive agenda is twofold: reduce upstream disparities that feed into decision-making, and reduce downstream harms caused by negative decisions or classifications. In practice this means not starting with the model but with the inequality; looking upstream at the structural conditions producing disparities; looking downstream at whether the stakes of a negative classification should be reduced; treating algorithms as auxiliary tools rather than as the solution itself; and using them, where appropriate, in service of broader reforms such as reducing access barriers, expunging records, supporting supportive rather than punitive interventions, or documenting systemic injustice (Green, 2021).

Application domains have repeatedly shown that marginal fairness does not settle the question. In recommender systems, work on Intersectional Two-sided Fairness Recommendation (ITFR) argues that a system may be fair on the user side and fair on the item side separately, yet still be unfair to specific Cartesian-product user-item groups. The motivating phenomenon is “diagonal unfairness”: for example, Male × Horror and Female × Children’s may be favored while Male × Children’s and Female × Horror are disadvantaged, despite acceptable user-side and item-side statistics (Wang et al., 2024).

Ranking research states the same point succinctly: fairness does not necessarily imply intersectionality. A ranking can satisfy fairness criteria on gender and on race separately, yet still exclude Black women or Hispanic men from the selected portion. The tutorial literature therefore organizes intersectional fair ranking through constraint-based methods, inference-model-based methods, and metrics-based methods, all aimed at representing joint, overlapping identities rather than one protected attribute at a time (Criscuolo et al., 7 Feb 2025).

Operational methods increasingly act directly on the joint subgroup space. FairHOME improves intersectional fairness at inference time by generating higher-order mutants that vary protected attributes across valid subgroup combinations, querying the same deployed model on the original input and its mutants, and ensembling the outputs. APFEx models intersectional fairness as a joint multi-objective optimization problem over the Cartesian product of sensitive attributes, while Mixed-Integer Optimization approaches enforce explicit subgroup fairness constraints and train intrinsically interpretable classifiers that can identify the most unfair subgroup exactly (Chen et al., 2024, Mondal et al., 17 Sep 2025, Němeček et al., 27 Jan 2026).

6. Normative disputes, governance, and open problems

One persistent dispute concerns whether fairness should be framed as equality or sufficiency. The hypothesis-testing literature argues that intersectional fairness should be understood less as “equal subgroup performance” than as a statistically cautious sufficiency standard: each group should receive enough, and fairness standards should not be lowered for smaller or historically disadvantaged groups. This directly challenges group-prevalence weighting schemes, which can make fairness constraints weaker as groups become rarer and can create perverse incentives against collecting more subgroup data (Himmelreich et al., 2024).

Another dispute concerns whether formal gains are substantively meaningful. The critique of leveling down argues that relative parity alone can conceal deterioration in group welfare. A model may look better under a relative criterion even when the best-performing group has merely been dragged downward, or when both best and worst groups are worse than before. The $VarRatio(K)$1-Intersectional Fairness proposal addresses this by combining a relative component with an absolute worst-group component, thereby making tradeoffs between parity and absolute performance explicit (Maheshwari et al., 2023).

Feminist accounts extend the critique from metrics to governance. The ROOF methodology—(R)ecognition, (O)vercoming, (O)vercoming, Ways (F)orward—proposes ten desiderata, including questioning the neutrality of decision processes, making positionality explicit, moving beyond the language of “intersectional algorithmic bias,” questioning the meaning of social categories, refusing to weigh or order oppression, mapping power and domination structures, acknowledging ripple effects, aligning purpose with context and impact, explicitly considering privileges and not only disadvantages, and recognizing that algorithmic systems may sometimes be used affirmatively rather than only defensively. A central implication is that principled non-deployment may be the most ethical option when an algorithm cannot be aligned with substantive justice aims (Mirsch et al., 25 Aug 2025).

Open problems remain extensive. Survey and methodological work repeatedly identify data sparsity, combinatorial explosion, missing or noisy demographic information, multiple testing, overlapping subgroup memberships, continuous protected variables, and the difficulty of choosing normatively acceptable thresholds such as a sufficiency level $VarRatio(K)$2 or an IJDI utility parameter $VarRatio(K)$3. The common conclusion is that metrics, by themselves, do not guarantee substantive justice; substantive intersectional fairness requires statistical caution, domain knowledge, governance, and attention to structural inequities beyond the model boundary (Gohar et al., 2023, Menghani et al., 2023).