- The paper argues that generative AI’s group representations should be evaluated by whether they support or undermine equal social standing, rather than by descriptive accuracy alone.
- The paper uses Fraser’s concept of participatory parity to show that accurate outputs can reinforce injustice, while group-targeted optimization may essentialize communities and suppress internal disagreement.
- The paper concludes that recognitional justice cannot be reduced to an algorithmic metric and requires public reasoning, meaningful participation, and collective contestation over AI-generated representations.
This paper argues that the dominant frameworks for evaluating how generative AI depicts social groups—grounded in descriptive accuracy and "representational fairness"—are structurally inadequate, and proposes replacing them with a standard of "recognitional justice" drawn from Nancy Fraser's theory of participatory parity (2608.12669). The argument is conceptual rather than empirical: the authors, both from Cornell's Department of Information Science, aim to reorient fair AI/ML and value alignment research away from epistemic criteria (truth, accuracy, fidelity to survey data) and toward a political criterion—whether AI outputs sustain or undermine the equal social standing of individuals and groups.
The distributive bias of fair AI
The paper begins by reconstructing the field's inherited division of labor. Fair AI/ML has long distinguished distributive (allocative) fairness—concerns about how algorithmic predictions mediate access to jobs, credit, parole, and treatment—from representational fairness, which concerns how systems characterize groups and accord social status (2608.12669). The distributive paradigm dominated because it fit the canonical application profile of predictive AI and because its problems, however difficult, admit quantitative formalization: demographic parity, equalized odds, and the broader "fairness toolbox" made harms specifiable, measurable, and optimizable (2608.12669). Representational harms, by contrast, are "intrinsically qualitative and thus resist from the outset standard computational abstractions" (2608.12669), which the authors identify as a principal reason the area remained comparatively underdeveloped.
The paper is nonetheless careful to credit existing work on representation, organizing it into three approaches: aggregate distributional analyses that apply a demographic-parity heuristic to label-group associations; instance-level harm taxonomies that deductively classify stereotyping and demeaning outputs; and analyses of omission and erasure, where groups are rendered invisible or dissolved into a dominant group's symbolic repertoire (2608.12669). The authors build on this work rather than dismissing it, but contend it remained secondary so long as high-stakes AI meant prediction and classification.
Generative AI as an expressive technology
The central empirical premise is that generative AI changes this balance. LLMs and text-to-image systems are "fundamentally expressive systems—their primary function is to convey meaning" that participates in the cultural construction of social groups (2608.12669). The paper identifies three interactive modes through which such systems generate value-laden group representations: personalization, in which models infer users' demographic, moral, and psychographic attributes; open-ended description of cultural communities, practices, and value orientations; and persona-based impersonation, in which models simulate members of groups as interactive stand-ins (2608.12669).
The implication the authors draw is that harms long catalogued in predictive contexts—biased associations, stereotyping, erasure—now extend across open-ended, multimodal, context-sensitive outputs, and that a single model can produce evaluative depictions of "any imaginable social group." Because framing choices alter which attributes are foregrounded and whether they are cast as virtues or deficits, representational flexibility is itself normatively significant (2608.12669).
The political theory interlude: distribution versus recognition
The paper then maps the parallel debate in political theory. Beginning in the 1990s, theorists including Young, Benhabib, Honneth, and Fraser argued that recognition is a dimension of justice irreducible to distribution (2608.12669). The paper distinguishes recognition as respect (universal equal moral standing, associated with Kantian universalism) from recognition as esteem (positive valuation of group-differentiated attributes, associated with multiculturalism and identity politics) (2608.12669). It surveys the reductive positions on both sides—the liberal "consumerist view" that recasts dignity and status as distributable goods, and Honneth's view that all injustice, including maldistribution, is ultimately misrecognition—and notes that the latter struggles with harms arising from impersonal economic structures, such as profit-driven layoffs (2608.12669).
The paper explicitly concedes it cannot resolve these debates, and that its argument "doesn't rest on the details of Fraser's account" (2608.12669). Fraser is chosen because her two-dimensional theory is a plausible integrative account that structurally parallels the fair AI literature's own distributive/representational split.
Participatory parity as the normative standard
Fraser's account holds that justice requires "parity of participation"—social arrangements permitting all adult members to interact as peers—satisfied by an objective condition (sufficient material resources) and an intersubjective condition (no institutionalized norms that systematically depreciate some categories of people) (2608.12669). Two features of this standard do the argumentative work in the paper.
First, it supplies a non-relativist test for recognition claims, evaluated at both intergroup and intragroup levels. Claims are warranted only where parity of participation is actually undermined, and proposed remedies are acceptable only if they do not themselves diminish it—for example, a minority religious group's claim to recognition for practices that subordinate women within the group must be rejected (2608.12669). Applied to AI, this yields a sharp thesis: representational inaccuracy is a diagnostic indicator of possible misrecognition, but not itself the injustice. An inaccurate depiction constitutes recognitional injustice only when it contributes to status subordination, between or within groups (2608.12669).
Second, participatory parity is deliberately non-computable. Fraser insists it "cannot be calculated by an algorithmic metric or method" and must be settled through public reasoning and contestation (2608.12669). This directly challenges the optimization template the paper identifies in cultural and pluralistic alignment research: construct a target dataset of group preferences or values (e.g., PRISM, GlobalOpinionQA), optimize models toward it via group-specific reward models or distributional preference optimization, and benchmark fidelity against empirical group-level distributions (2608.12669).
Three failures of accuracy-oriented remedies
The paper's core contribution is a systematic critique of this pipeline, organized as three distinct failure modes.
Accuracy can entrench injustice. Drawing on Noble's analysis of image search—where "accurate" gender-skewed results for doctors and nurses nonetheless reinforced occupational status hierarchies—the paper argues that faithful representations of an unjust social reality can perpetuate that reality (2608.12669). This raises a question accuracy frameworks cannot answer: should models reflect existing conditions or deliberately depart from them, and where does justified departure become counterproductive distortion? The authors note, candidly, that the case for "inaccurate but recognitionally reparative" outputs must also contend with trade-offs against reliability and trust, and that the matter "would need to be decided collectively" (2608.12669). The paper draws a methodological parallel to Green's critique of formal algorithmic fairness: optimizing against a technical target cannot by itself establish that an intervention is just (2608.12669).
Authority over representation is politically constituted. Even granting that accuracy is the right goal, someone must decide what accurately represents a group. Participatory approaches partially address this, but the paper stresses—citing critiques of participatory AI—that participation is a spectrum from tokenism to real influence, with developers retaining control over participant selection, deliberative scope, and whether input changes the model (2608.12669). The paper's position here is measured: this limitation "does not make participation irrelevant," since participation's value lies partly in exposing the political choices hidden behind claims of accuracy (2608.12669).
Groups cannot be stabilized as optimization targets. The deepest problem is ontological. Social groups are hybridized, internally plural, and continuously contested; Benhabib's characterization of culture as a polyvocal, intergenerational conversation forecloses any neutral referent for accuracy (2608.12669). The paper argues that fixing a group as a target for optimization risks reproducing the essentialism such remedies aim to correct—statistical predictors privilege dominant, frequent expressions of cultural identity, and even participatory RLHF aggregates divergent annotator judgments into a single scalar reward signal, compressing intragroup disagreement into one average target (2608.12669). Formalization forces choices about whether disagreement is averaged, weighted, partitioned, or preserved across multiple models, and each choice elevates some interpretations while marginalizing others (2608.12669).
Limitations and open questions
The paper is a position paper, and its limitations are largely inherent to that genre. It offers no operational procedure for applying participatory parity to concrete AI systems; indeed, its thesis depends on the claim that no such procedure is possible, which leaves practitioners without actionable guidance beyond "public contestation." The normative framework is asserted as plausible rather than defended against rival political theories—the paper explicitly brackets relational egalitarianism and related accounts (2608.12669). The trade-offs between recognitionally reparative but descriptively inaccurate outputs and values such as reliability and trust are acknowledged but not analyzed (2608.12669). And the argument's reliance on Fraser's intersubjective condition presupposes that "institutionalized patterns of cultural value" can be identified in AI systems at a level of abstraction that supports collective judgment—a presupposition the paper does not test empirically. Open questions the paper leaves include: what institutional forms of public contestation over model outputs would satisfy the discursive requirement, and how quality-of-service disparities, which the paper shows entangle recognition with redistribution, should be adjudicated within a two-dimensional frame (2608.12669).
Conclusion
The paper's claim, stated compactly: the deep question about generative AI's depictions of social groups is not whether they are accurate but whether they strengthen or undermine parity of participation, and answering it "cannot be computed or decided by experts or engineers behind closed doors" (2608.12669). For researchers in fair AI/ML and value alignment, the practical upshot is a reordering of priorities—accuracy metrics and participatory data collection remain diagnostically and politically useful, but neither constitutes justice. What the paper establishes is a conceptual redirection; what it leaves undetermined is the institutional machinery through which the required collective deliberation would actually govern model development.