From Fair Representation to Just Recognition in Generative AI

AI summary

Overview

Research area: AI ethics and normative political theory — specifically the intersection of the fair AI/ML literature on "representational fairness," the value-alignment literature on what and whose values AI systems should represent, and theories of "justice as recognition" from political philosophy.

Technical level: Beginner-Friendly. The paper is a conceptual and normative argument. It contains no models, datasets, benchmarks, metrics, or experiments, and no mathematics or code.

Scope: A single-sentence scope: the paper argues that representational harms in generative AI should be diagnosed not as problems of misrepresentation (an epistemic, accuracy-based frame) but as problems of misrecognition (a political, status-based frame), using Nancy Fraser's "two-dimensional" theory of justice and its standard of "parity of participation" as the evaluative tool.

What This Paper Is About

Fair AI research has traditionally split AI harms into distributive (allocative) fairness — how algorithms allocate resources, goods, and opportunities — and representational fairness — how AI shapes the way people and social groups are perceived, understood, and accorded social status. The authors argue that generative AI, being fundamentally expressive rather than predictive, makes representational issues urgent, yet the field's dominant response — judging outputs by whether they accurately depict social groups — is inadequate because group boundaries are contested, no one has clear authority to certify accuracy, and even accurate representations can reinforce harmful hierarchies. Their goal is to justify a conceptual shift from "representational fairness" to "recognitional justice," drawing on political theory.

Key Contributions

1. A diagnosis of why accuracy-based evaluation fails for generative AI. The authors identify three specific shortcomings of judging AI representations against standards of descriptive accuracy: for many (if not most) social groups there are no stable or bounded referents against which to judge representational inaccuracy; it is unclear who has the authority to decide what counts as misrepresenting a particular group; and even accurate representations can reproduce harmful patterns detrimental to group outcomes.

2. A reframing of representational harm as misrecognition rather than misrepresentation. Rather than asking whether AI depictions are accurate, the paper proposes asking whether they are just — treating the harm of stereotypical or demeaning AI output as the diminishment of people's ability to participate in society as equals.

3. Transposition of Nancy Fraser's two-dimensional theory of justice into AI ethics. The paper introduces Fraser's "parity of participation" and its two conditions (an objective, material-resource condition and an intersubjective condition barring institutionalized norms that systematically depreciate categories of people) as an evaluative standard for generative AI outputs, including its two-level test at the intergroup and intragroup levels.

4. An inventory and reinterpretation of generative AI's three modes of representation. The paper characterizes generative AI as representing users (inferring demographics, intentions and goals, moral commitments, and psychographic and behavioral profiles), producing evaluative depictions of social groups and cultural symbols across contexts, and impersonating group members through personas — and maps existing representational-fairness harms onto these modes.

Main Findings

• The field's distributive bias is a matter of fit, not just neglect. The authors report that distributive fairness produced a conceptual and technical "fairness toolbox" — formal criteria such as demographic parity and equalized odds that made harms specifiable, measurable, and optimizable in specific decision domains — whereas research on representational fairness has been a "relative backwater." They attribute this to the distributive frame fitting the majority of real-world prediction cases (hiring, credit scoring, recidivism, medicine) and to the fact that questions about representation are intrinsically qualitative and resist standard computational abstractions.

• Three existing approaches to representational harm in fair AI. The paper identifies: (1) an aggregate, distributive logic applying a demographic parity heuristic (research on search engine image results finding women depicted as "smiling" or "beautiful" while men are characterized as "leaders" or "engineers"); (2) taxonomies of harm specified at the level of individual evaluative instances (for example, assigning an animal label to an image of a person of a particular race); and (3) absence of representation, which occurs either as failure to represent a group at all or as erasure — assigning a group the symbolic representations of another, typically dominant, cultural group (for example, the lack of image labels for nonbinary gender identities).

• Generative AI widens the scope of representational concern. Unlike predictive systems, which represent individuals within contextually bounded, task-specific frames such as creditworthiness, employability, or risk of reoffending, a single generative model can produce numerous types of evaluations of any imaginable social group, across open-ended, multimodal, and context-sensitive representations. A change in conceptual or stylistic framing can alter which attributes are foregrounded, whether they are cast as virtues or deficits, and what stance toward the group the description invites.

• Recognition is a distinct dimension of justice in political theory. The paper distinguishes two dimensions commonly discussed: recognition as respect, grounded in the universal and equal standing of all persons by virtue of their common humanity, expressed politically in a politics of universalism; and recognition as esteem, concerning positive valuation of the particular features through which individuals and groups understand themselves, associated with communitarianism, multiculturalism, and identity politics.

• Fraser's nonreductive alternative. Fraser argues that recognition and distribution are both basic, co-constitutive components of social justice, and that the central demand of justice is equal social standing, or parity of participation — social arrangements that permit all (adult) members of society to interact with one another as peers. Her account supplies two conditions: an objective condition requiring sufficient material resources to interact as peers, and an intersubjective condition precluding institutionalized norms that systematically depreciate some categories of people and the qualities associated with them, either by burdening them with excessive ascribed "difference" or by failing to acknowledge their distinctiveness. Illustrative cases include a black Wall Street banker who finds it difficult to hail a taxi; women disadvantaged both through underpaid labor and through sexually objectifying cultural representations; and marginalized racial groups overrepresented in low paid work while also subjected to value schemes that privilege traits associated with whiteness.

• A two-level test that avoids relativism. At the intergroup level, inequalities must be eliminated when one group's standing is systematically lowered relative to another (the paper's example: marriage laws excluding same-sex partnerships as illegitimate). At the intragroup level, claims for recognition must be assessed according to whether the practices for which recognition is sought themselves preserve parity of participation (the paper's example: a minority religious group seeking recognition for practices that undermined women's equal standing). Claims advanced by white supremacist groups that their self-esteem is compromised by living alongside "inferior" races are unjustified and warrant no remedy.

• Accuracy is neither necessary nor sufficient for justice. An inaccurate group representation may constitute recognitional injustice when it contributes to status subordination — that is, when it undermines social arrangements permitting all adult members of society to interact as peers — either between groups (for example, between a majority and a minority) or within a group. The paper states that representational inaccuracy alone does not demonstrate injustice.

• Recognition and distribution are entangled in concrete generative AI cases. The paper notes that disparities in quality of service (for example, voice-recognition systems with lower accuracy rates for underrepresented languages and culturally marginalized users) can simultaneously restrict access to valuable technological resources and reproduce the diminished social standing of disadvantaged groups. A rich literature in AI fairness has documented cases of disparate quality of service.

• Judgments about participatory parity cannot be made monologically. There is no objective marker that signals when parity of participation has been achieved, because the meaning of any such marker is itself open to interpretation and contestation. Judgments must be worked out discursively and dialogically through public reasoning and contestation; what constitutes just recognition cannot be determined by designers or technical procedures alone.

• No empirical results are reported. The paper presents no experiments, datasets, benchmarks, or quantitative findings.

Methodology in Plain English

The paper is a conceptual and normative argument rather than an empirical study. The authors proceed by literature review and theoretical comparison across two fields: they survey the fair AI/ML distinction between distributive and representational fairness (including taxonomies and definitions of representational harm), and they survey value-alignment research on pluralistic alignment — work such as collecting preference data from heterogeneous populations, distributional parity measures, benchmarks approximating pluralistic value datasets, and comparisons of model outputs to empirically observed value distributions from the World Values Survey. They then review debates in political theory between justice as redistribution and justice as recognition, including arguments that reduce one to the other and arguments against such reduction. From these debates they select Nancy Fraser's two-dimensional theory of justice and its notion of parity of participation as the framework to carry over, explain its two conditions and its two-level test of recognition claims, and then re-describe generative AI's three modes of interaction (personalization, evaluative description of social groups and cultural symbols, and persona impersonation) in its terms. The paper explicitly states that its overarching argument does not rest on the details of Fraser's account, and that it does not resolve the underlying debates in political theory.

Why This Matters

Impact on research: The paper offers AI ethics communities — fair AI/ML and value alignment in particular — an alternative normative vocabulary for a class of harms the authors say the field lacks adequate tools to diagnose and address. It reframes pluralistic alignment questions (what and whose values systems should represent) as questions that cannot be settled by epistemic criteria of truth and accuracy, and it points to a nonrelativist standard that can distinguish justified from unjustified claims of representational harm.

Real-world applications (drawn from the paper's own examples):

• Conversational systems that produce value-laden depictions of social groups, cultural symbols, practices, and roles — for example, explanations of the ritual practices of a religious holiday in another country, or accounts of dating expectations in two different cultures.
• Image generation of a "traditional" wedding, a "safe" neighborhood, or a "professional" woman.
• Text-to-image systems exhibiting exoticism and cultural misappropriation, where cultures are represented through homogenized, misplaced, or otherwise culturally inaccurate details.
• Persona-based systems used as intimate companions, friends, romantic partners, mentors, teachers, and health advisors, plus culturally aligned helper models that rate how well candidate responses represent a target culture.
• Voice-recognition systems with lower accuracy rates for underrepresented languages and culturally marginalized users.

Industry relevance: The argument bears on practitioners who build and evaluate generative systems through persona prompting, preference-data collection from social groups, and comparison of outputs to survey data. The authors argue these methods, while valuable, are incomplete on their own, and that decisions about just recognition must ultimately be defined, debated, and contested collectively in public — a constraint on claims that such questions can be resolved by designers or technical procedures alone.

Future Directions

• Develop non-epistemic evaluation practices for generative AI outputs. If justice rather than accuracy is the standard, the field needs ways of assessing whether a representation contributes to status subordination between or within groups, and this remains an open problem.
• Work out how participatory parity is assessed dialogically. The paper insists judgments must be made through public reasoning and contestation but does not specify institutional or procedural mechanisms for doing so at the scale of deployed generative models.
• Reconcile value-alignment practices with recognitional justice. The paper calls for deeper engagement between pluralistic alignment methods (preference collection, distributional parity measures, pluralistic value benchmarks) and the normative questions those methods implicitly answer.
• Untangle intertwined distributive and recognitional harms. Cases such as quality-of-service disparities show the two dimensions are rarely neatly delineated in practice; the paper leaves open how remedies should be tailored when both are at stake.

Target Audience

AI ethics and fairness researchers, especially those working on representational harm, bias evaluation, and pluralistic value alignment; value-alignment and AI safety practitioners who design evaluations of model outputs about social groups; policy analysts and advocates concerned with how generative systems depict cultures, identities, and communities; and political philosophers or social theorists interested in how recognition theory applies to machine-generated expression. The paper is accessible without technical background, since it contains no formal methods or empirical results.

Authors’ abstract

The fair AI/ML literature has long distinguished distributive fairness, concerning how automated systems allocate resources and opportunities, from representational fairness, concerning how they shape the ways individuals and social groups are perceived, understood, and accorded social status. Generative AI is rebalancing these normative dimensions. Unlike predictive systems, large language models (LLMs) and related technologies are fundamentally expressive: their primary function is to convey meaning rather than automate domain-specific decisions. Representational harm has also become central to value alignment, especially in research on what and whose values and perspectives AI systems should represent. Existing approaches to harms in the representation of social groups often appeal to descriptive accuracy, but this strategy has important limitations. For many social groups, no stable or bounded referent exists against which representational accuracy can be judged. It is also unclear who has the authority to decide what counts as misrepresentation, while even accurate representations can reproduce harmful social patterns. The underlying problem, we argue, is therefore not simply misrepresentation but misrecognition. Drawing on political theory, especially Nancy Fraser's account of participatory parity, we show how moving from representational fairness to recognitional justice provides better conceptual and normative tools for governing central fairness challenges in generative AI.