arxiv: 2604.22675 · v1 · submitted 2026-04-24 · 💻 cs.SI

Recognition: unknown

Measuring Epistemic Unfairness for Algorithmic Decision-Making

Camilla Quaresmini , Lisa Piccinin , Valentina Breschi

Authors on Pith no claims yet

Pith reviewed 2026-05-08 09:06 UTC · model grok-4.3

classification 💻 cs.SI

keywords epistemic injusticealgorithmic fairnessepistemic unfairnessdeficit measurementrecommender systemsopinion dynamicsfairness auditingcapability inequity

0 comments

The pith

A deficit-based framework quantifies epistemic injustice in algorithms as gaps in credibility, uptake, and agency that persist even when predictive fairness holds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a measurement system for harms that affect what people can know and believe, rather than just how accurately an algorithm predicts outcomes. It models these harms as shortfalls between ideal and actual conditions for credibility, uptake, and epistemic agency, then maps the shortfalls onto specific stages where algorithms filter or present information. This approach shows that standard checks on error rates or group parity can leave epistemic problems untouched. By adapting inequality indices to either direct distributions of epistemic goods or to the opportunities they create, the framework enables repeated checks as systems run over time. A simulation of opinion dynamics under recommender interventions illustrates how the indices track the growth or reduction of these problems.

Core claim

We propose a quantitative framework for evaluating forms of epistemic injustice in algorithmic environments. First, we introduce a deficit-based template that models epistemic injustices as gaps between ideal and realized conditions across features such as credibility, uptake, and epistemic agency. We map these deficits to concrete stages of algorithmic mediation, showing how epistemic injustice can persist even when standard fairness constraints are satisfied. Drawing on distributive fairness indices, we distinguish two evaluation stances: resource inequality, where indices are applied to distributions of epistemic goods directly, and capability/rights inequity, where indices are applied to

What carries the argument

A deficit-based template that treats epistemic injustices as measurable gaps in credibility, uptake, and epistemic agency, mapped onto algorithmic mediation stages, then evaluated with translated distributive fairness indices in either resource or capability stances.

If this is right

Epistemic unfairness can be detected and tracked separately from standard predictive fairness metrics.
The indices identify distinct patterns such as exclusionary tails and hierarchical concentration in access to epistemic goods.
Audits can be performed longitudinally as algorithms are updated iteratively.
In recommender-mediated settings the indices register changes in epistemic unfairness driven by repeated platform interventions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same deficit template could be tested on non-recommender systems such as search or content moderation to check whether epistemic gaps appear in those mediation stages as well.
If the indices are deployed in live systems, one could measure whether changes in the indices precede observable shifts in user belief diversity or participation rates.
Regulators might require these epistemic measures alongside accuracy audits when approving high-impact algorithmic tools.

Load-bearing premise

Epistemic ideas like credibility and agency can be turned into precise numerical deficits that fit algorithmic steps without losing their original meaning or creating unresolvable measurement conflicts.

What would settle it

Run the recommender opinion-dynamics simulation with balanced predictive fairness constraints; if the epistemic indices stay near zero while users show reduced access to diverse information or lowered willingness to share beliefs, the framework fails to capture the intended harms.

Figures

Figures reproduced from arXiv: 2604.22675 by Camilla Quaresmini, Lisa Piccinin, Valentina Breschi.

**Figure 1.** Figure 1: The proposed framework for evaluating algorithmic epistemic injustice. The gap between the normative Epistemic Ideal (I ∗ ) and the Realized Condition (I) is modeled as a measurable deficit (∆I ). and attention are scarce and competitively allocated, an unearned credibility excess for dominant users can directly generate a credibility deficit for marginalized ones. While the current framework prioritizes m… view at source ↗

**Figure 2.** Figure 2: Conceptual pipeline underlying our framework. The figure summarizes the diagnostic structure of the paper. promotion satisfies parity at one step, but also whether the feedback loop concentrates epistemic authority, produces lower-tail deprivation, or generates diverging epistemic standing within and across groups. These aspects are however not assessed by mainstream fairness metrics, that are mostly parti… view at source ↗

**Figure 3.** Figure 3: Simulation framework. Network structure and distribution of initial opinions for the two groups. Group A information is depicted in red, while Group B is indicated in blue. excluded. For an output-oriented analysis, let νi denote output-induced epistemic opportunity (e.g., expected uptake or exposure that contributes to realized epistemic capital over time). The same indices summarize how evenly the system… view at source ↗

**Figure 4.** Figure 4: Intervention effects on fairness indices. Analysis on input and output quantities comparing the three presented scenarios. (amplification is applied to a randomly selected subset of N/2 agents). At intervention times, the platform amplifies selected influence links and then renormalizes the network to preserve its structural properties (see Appendix A for simulation details). Fig. 4a shows that both interv… view at source ↗

read the original abstract

Algorithmic systems increasingly function as epistemic infrastructures that govern the conditions of interpretative access and social belief. Yet, mainstream auditing strategies operationalize fairness primarily in predictive terms - error rates, calibration, or group-level parity - leaving epistemic harms under-theorized and under-measured. We propose a quantitative framework for evaluating forms of epistemic injustice in algorithmic environments. First, we introduce a deficit-based template that models epistemic injustices as gaps between ideal and realized conditions across features such as credibility, uptake, and epistemic agency. We map these deficits to concrete stages of algorithmic mediation, showing how epistemic injustice can persist even when standard fairness constraints are satisfied. Drawing on distributive fairness indices, we distinguish two evaluation stances: resource inequality, where indices are applied to distributions of epistemic goods directly, and capability/rights inequity, where indices are applied to output-induced epistemic opportunity. We provide an epistemic translation of canonical indices, illustrating how they diagnose complementary signatures of unfairness - such as exclusionary tails and hierarchical concentration - and support longitudinal auditing under iterative deployment. We also provide a simulation study of a recommender-mediated opinion dynamics setting, showing how the proposed indices capture the evolution of epistemic unfairness under repeated platform interventions. The result is a measurement framework that makes the epistemic dimension of algorithmic harms explicit for system design and evaluation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper adapts distributive fairness indices into a deficit template for epistemic harms like credibility gaps in algorithms and includes a recommender simulation, but does not show these measures are independent of standard predictive fairness.

read the letter

The core contribution is the deficit-based template that maps epistemic features such as credibility, uptake, and agency to stages of algorithmic mediation, then translates canonical fairness indices into resource-inequality and capability-inequity versions. The simulation of opinion dynamics under repeated platform interventions gives a practical illustration of how these indices might track changes over time. That combination is new relative to the predictive-fairness literature referenced in the abstract, and the resource-versus-capability split is a reasonable way to organize the evaluation stances. The longitudinal angle in the simulation is also a step toward usable auditing tools for iterative systems like recommenders. Those elements are the parts that hold up on a first read. The main limitation is that the independence claim is not backed by an explicit check. The abstract states that epistemic injustice can persist even when standard fairness constraints are satisfied, yet nothing described shows the epistemic indices are mathematically or empirically separate from the data patterns already captured by parity or calibration metrics. If platform changes that satisfy demographic parity also drive the deficit scores to zero, the persistence result reduces to a restatement rather than a new dimension. The formal derivation of the index translations is not detailed enough to judge whether the mapping avoids circularity when ideal conditions are defined in terms of the same algorithmic outputs. Validation against external epistemic benchmarks is also missing. This paper is for researchers working on fairness auditing in social platforms and AI ethics who want to extend beyond error rates. A reader building evaluation frameworks could borrow the template and simulation design as a starting point. I would send it for peer review because the proposal is coherent and the simulation provides some grounding, even though the independence argument needs tightening.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a quantitative framework for evaluating epistemic injustice in algorithmic systems. It introduces a deficit-based template modeling gaps in credibility, uptake, and epistemic agency mapped to stages of algorithmic mediation, adapts distributive fairness indices under resource-inequality and capability-inequity stances, and presents a simulation of recommender-mediated opinion dynamics to illustrate that epistemic unfairness can persist even when standard predictive fairness constraints are satisfied.

Significance. If the indices are shown to be independent of standard fairness metrics, the work would provide a valuable extension of auditing tools to epistemic harms such as exclusionary tails and hierarchical concentration in belief formation. The simulation study offers a concrete demonstration in an iterative setting, and the use of established indices supports longitudinal evaluation. These elements position the framework as a practical addition for system design beyond predictive parity.

major comments (2)

The central claim that epistemic deficits persist independently of standard fairness metrics lacks an explicit orthogonality check. The recommender opinion-dynamics simulation does not report results for interventions that satisfy demographic parity or equalized odds and whether those drive the proposed epistemic indices to zero; without this, the persistence result risks reducing to a restatement of existing fairness violations rather than identifying a new dimension.
The deficit-based template and its epistemic translation of canonical indices are presented conceptually without formal equations or derivation steps showing how credibility/uptake/agency gaps are quantified from algorithmic outputs. This leaves the mapping open to post-hoc fitting concerns and makes it difficult to verify that the indices capture variation not already present in the data distributions used by standard metrics.

minor comments (2)

The abstract states that the framework supports longitudinal auditing under iterative deployment, yet the simulation results section does not include time-series plots or analysis of index evolution across repeated platform interventions.
Notation for the translated indices (e.g., how resource vs. capability versions differ mathematically) would benefit from explicit definitions or pseudocode rather than purely descriptive text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which helps clarify the independence of the epistemic framework and strengthens its formal presentation. We address each major comment below and commit to revisions that directly incorporate the suggested checks and formalizations.

read point-by-point responses

Referee: The central claim that epistemic deficits persist independently of standard fairness metrics lacks an explicit orthogonality check. The recommender opinion-dynamics simulation does not report results for interventions that satisfy demographic parity or equalized odds and whether those drive the proposed epistemic indices to zero; without this, the persistence result risks reducing to a restatement of existing fairness violations rather than identifying a new dimension.

Authors: We agree that an explicit orthogonality demonstration is necessary to establish the framework as capturing a distinct dimension. In the revised manuscript, we will extend the simulation section with new experiments that enforce demographic parity and equalized odds constraints on the recommender interventions. We will report the corresponding epistemic index trajectories (under both resource-inequality and capability-inequity stances) and show that they remain strictly positive, with quantitative tables and plots comparing them to the unconstrained case. This will directly address the concern that the persistence result might collapse to standard fairness violations. revision: yes
Referee: The deficit-based template and its epistemic translation of canonical indices are presented conceptually without formal equations or derivation steps showing how credibility/uptake/agency gaps are quantified from algorithmic outputs. This leaves the mapping open to post-hoc fitting concerns and makes it difficult to verify that the indices capture variation not already present in the data distributions used by standard metrics.

Authors: We acknowledge that the current conceptual presentation would benefit from explicit formalization to enable verification and reduce ambiguity. The revised manuscript will add a new subsection (Section 3.1) containing: (i) formal definitions of the deficit template, e.g., credibility deficit as the expected difference between an agent's ground-truth expertise-based credibility score and the algorithmically assigned score derived from output logits or rankings; (ii) step-by-step derivations translating the Gini coefficient and Theil index to epistemic goods under both evaluation stances, with explicit formulas showing how gaps in uptake and agency are extracted from algorithmic outputs; and (iii) a brief variance decomposition argument demonstrating that the epistemic indices are not linear functions of the input feature distributions used by standard metrics. These additions will allow readers to replicate the mapping and confirm the captured variation. revision: yes

Circularity Check

0 steps flagged

No circularity: framework extends canonical indices via explicit translation and simulation

full rationale

The derivation introduces a deficit template mapping epistemic concepts to algorithmic stages and translates existing distributive fairness indices into resource-inequality and capability-inequity stances. These steps rely on external canonical indices rather than self-defining the target quantities or fitting parameters to the outputs being audited. The simulation demonstrates evolution under interventions without reducing the epistemic signatures to the same predictive fairness metrics by construction. No self-citation chains, uniqueness theorems, or ansatzes imported from prior author work appear as load-bearing elements. The central claim of complementary signatures is therefore independent of the inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the assumption that epistemic injustice admits a deficit-based quantitative representation that can be applied stage-wise to algorithmic processes; no free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption Epistemic injustices can be modeled as gaps between ideal and realized conditions in credibility, uptake, and epistemic agency.
This is the foundational template invoked to operationalize the measurement framework.

pith-pipeline@v0.9.0 · 5533 in / 1192 out tokens · 55996 ms · 2026-05-08T09:06:07.811155+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

[1]

What are you doing, TikTok?

Epistemic injustices and curriculum: Strategizing for justice.Social Sciences & Humanities Open11 (2025), 101220. [Baker and Hawn(2022)] Ryan S Baker and Aaron Hawn. 2022. Algorithmic bias in education.International journal of artificial intelligence in education32, 4 (2022), 1052–1092. [Barocas et al.(2020)] Solon Barocas, Moritz Hardt, and Arvind Naraya...

work page arXiv 2025
[2]

V oiceless

On the information unfairness of social networks. InProceedings of the 2020 SIAM International Conference on Data Mining. SIAM, 613–521. [Keyes(2018)] Os Keyes. 2018. The misgendering machines: Trans/HCI implications of automatic gender recognition. Proceedings of the ACM on human-computer interaction2, CSCW (2018), 1–22. [Kim(2022)] Junyeol Kim. 2022. De...

work page arXiv 2020
[3]

In2nd European Workshop on Algorithmic Fairness, EWAF 2023

Qualification and quantification of fairness for sustainable mobility policies. In2nd European Workshop on Algorithmic Fairness, EWAF 2023. CEUR-WS. org, 29. [Quaresmini et al.(2025)] Camilla Quaresmini, Eugenia Villa, Samira Maghool, Valentina Breschi, Viola Schaffonati, and Mara Tanelli. 2025. The Role of Epistemic Fairness in Dynamics Models to Support...

work page doi:10.1016/j.ifacol.2025.12.026 2023