Community-Defined AI Value Pluralism

• CDAVP is a socio-technical framework that redefines AI governance by empowering communities to create rich, machine-readable value profiles.
• It enables user-controlled, context-sensitive activation of community-defined value profiles to adapt system behavior to diverse environments.
• The framework balances pluralism with bound meta-rules, ensuring legitimacy, auditability, and proactive conflict moderation.

Searching arXiv for the primary CDAVP paper and closely related pluralism/alignment work to ground the article in current literature. Community-Defined AI Value Pluralism (CDAVP) is a “socio-technical framework” proposed for AI governance under conditions of durable value disagreement, heterogeneous communities, and contested legitimacy. It is introduced as a response to a “Value Alignment Gap”: the mismatch between the aspiration to align AI with “human values” and the practical impossibility of defining one legitimate value system for all people and contexts. Rather than seeking a single globally “aligned” model, CDAVP “reframes the design problem from achieving a single ‘aligned’ state to infrastructuring a dynamic ecosystem for value deliberation and application.” In that ecosystem, communities define “rich, machine-readable representations” of shared values, users retain “ultimate control (agency)” over which profiles are active in context, and all applications remain bounded by “non-negotiable, democratically-legitimated meta-rules” (Mayer, 7 Jul 2025).

1. Conceptual core and problem setting

CDAVP begins from the claim that prevailing alignment practice is too centralized and too static. In the framework’s own diagnosis, current approaches rely on top-down value definition by firms, designers, regulators, or technical researchers, leaving “users and communities unable to challenge or shape the values embedded in the systems that govern their digital lives.” The result is described as a “crisis of legitimacy and trust,” together with a broader “crisis of contestability in an AI-authored world” (Mayer, 7 Jul 2025).

The framework’s central normative move is to relocate value definition away from universalized system-wide defaults and toward communities and users. “Community” is not treated as a mere aggregate of users or a passive population of data subjects. In related work on community co-creation, it is defined as “a group of people who are indirectly or directly affected by issues in civil society and are dedicated to making sure that these issues are recognized and resolved,” while “community empowerment” is “a process of yielding agency to communities so that they can use technology, data, and informed rhetoric to create and disseminate evidence to advocate for social and policy changes” (Hsu et al., 2021). CDAVP generalizes that orientation from local civic AI projects to a wider architecture of value governance.

The framework also rejects the idea that AI pluralism is exhausted by output variation or post hoc personalization. A closely related critique argues that AI systems do not only represent diverse values or preferences; they also impose ontologies, determining “what counts as an entity, relation, feature, harm, benefit, and valid form of evidence.” On that view, pluralism limited to outputs remains incomplete unless affected actors have “procedural standing” over problem framing, categories, evidence, aggregation, evaluation, revision, and appeal (Mushkani, 15 Jun 2026). CDAVP is therefore best understood as a claim about normative authority, not merely about user customization.

2. Intellectual context in pluralistic alignment and sociotechnical governance

CDAVP belongs to a broader shift from monolithic alignment to sociotechnical alignment and normative pluralism. Work on AI governance has argued that legitimacy depends “not only on technical correctness but on alignment with societal norms, democratic principles, and human dignity,” and that “public values” should be treated as “operational imperatives that determine AI legitimacy and social license to operate” (Nkongolo, 4 Feb 2026). In that literature, value pluralism is not a defect to be averaged away but a constitutive condition of governance.

A parallel line of work in full-stack alignment argues that “beneficial societal outcomes cannot be guaranteed by aligning individual AI systems with the intentions of their operators or users,” because institutions and markets can distort values as they move upward through sociotechnical layers. It proposes “full-stack alignment” as “the robust co-alignment of AI systems and institutions with what people value,” together with “thick models of value” that preserve “justifications, social meanings, and constitutive relationships” rather than collapsing everything into preferences or unconstrained text (Edelman et al., 3 Dec 2025). CDAVP is compatible with that diagnosis, but it gives the community a more explicit role in defining and maintaining the value representations themselves.

Within pluralistic alignment research on LLMs, three model-level notions have been formalized: “Overton pluralistic models” that present a spectrum of reasonable responses, “steerably pluralistic models” that can steer to reflect certain perspectives, and “distributionally pluralistic models” that are calibrated to a target population (Sorensen et al., 2024). Those distinctions are important for CDAVP because they separate three different tasks: exposing legitimate disagreement, conditioning on a specified community profile, and reproducing within-community heterogeneity.

A further technical strand studies whether a society should be represented by a single value model at all. One proposal formalizes the “value system of a society” as “a set of value systems of different groups, rather than as the simple aggregation of individual value systems” (Holgado-Sánchez et al., 28 Jul 2025). That result supports pluralism at the group level, but its “groups” are inferred by clustering rather than self-defined through participatory governance. CDAVP diverges precisely at that point: it is not only pluralistic, but explicitly community-defined.

3. Architectural structure: profiles, activation, and bounded pluralism

The CDAVP framework is organized around “three interdependent pillars” within an overarching layer of meta-rules (Mayer, 7 Jul 2025).

The first pillar is the community-defined value profile. A value profile is a “rich, structured, and machine-readable representation of a community’s shared values.” The framework explicitly states that these profiles are not simple preference lists: they may include “not only preferences but also community-specific rights and duties.” Illustrative fragments in the framework include deontic rules such as “RULE: The process to cancel a subscription must not require more steps than the sign-up process” and weighted institutional commitments such as Priority_Weight(Violent_Crime) = 10 and Priority_Weight(Minor_Infractions) = 1 (Mayer, 7 Jul 2025). The designer’s role is therefore shifted away from fixing one static normative interface and toward building infrastructures through which communities can define, revise, and contest explicit machine-readable norms.

The second pillar is user-controlled, context-sensitive activation. CDAVP assumes that users often belong to multiple communities and that values are context dependent. For that reason, the user “retains ultimate control (agency)” over which profiles are active in a given situation. The framework’s examples include selecting a “Scientific Discourse” profile for one context and a “Family-Friendly” profile for another. This arrangement distinguishes collective articulation from individual activation: the normative content of a profile is collective, but the choice of which profile should govern a particular interaction belongs to the user (Mayer, 7 Jul 2025).

The third pillar is systemic application and privacy-preserving conflict moderation. AI applications are expected to interpret the active profiles and adapt behavior accordingly, while platforms moderate conflicts “transparently” and in a “privacy-preserving” manner. The paper does not provide a universal runtime semantics or conflict-resolution algorithm, but it is explicit that conflicts may arise either among multiple profiles activated by one user or across users in shared spaces (Mayer, 7 Jul 2025). In that sense, CDAVP is an architectural proposal rather than a single optimization objective.

All three pillars are bounded by “non-negotiable, democratically-legitimated meta-rules.” These are the framework’s constitutional layer: universal human rights, established law, prevention of direct harm, protection of minors, and ecological sustainability are cited as examples. CDAVP is therefore not an unrestricted relativism. Its pluralism is explicitly bounded.

4. Contestability, evaluation, and pluralism as an auditable property

CDAVP operationalizes contestability at multiple levels. Communities can define and revise profiles; subgroups can fork and create new profiles; users can select, activate, switch, and combine profiles; institutions can be required to publish explicit value profiles; and platforms can be audited for fairness, transparency, and compliance (Mayer, 7 Jul 2025). This multi-level structure is one of the framework’s most distinctive features: contestability is not limited to appealing an output after the fact, but includes contesting the value basis of system behavior in advance.

That emphasis connects directly to measurement and evaluation research. Work on “ValueCompass” provides a framework of 49 “fundamental values,” grouped into 12 motivational value types and 5 higher-order dimensions, together with a vignette-based “Value Form” for asking respondents “To what extent do you agree or disagree that AI should…” embody those values. The key empirical result is that values differ across healthcare, education, public sector, and collaborative writing scenarios, supporting “context-aware AI alignment strategies” rather than a uniform value target (Shen et al., 2024). For CDAVP, that provides a measurement substrate for community-specific value elicitation and audit.

Other work places community authority earlier in the measurement pipeline. In research on AI-generated cultural artifacts, community involvement is concentrated in the “systematizing” stage of measurement: communities define what “cultural appropriateness” means before technical teams operationalize that definition into a rubric or automated judge. The paper’s central lesson is that “community participation materially improves what is being measured,” but that downstream automation remains partial and validity-sensitive (Johnson et al., 2 Apr 2026). CDAVP can be read as extending that logic beyond evaluation into the entire lifecycle of model behavior.

Pluralism can also be audited behaviorally. In moderation research, disagreement among perspective-differentiated agents is treated not as noise but as signal: “disagreement structure—not magnitude—guides when human judgment is needed,” and “convergent disagreement” is proposed as a marker of genuine value pluralism rather than mere error (Wawer et al., 4 Apr 2026). In clinical ethics, frontier models are shown to exhibit “Overton pluralism” in their reasoning while remaining “near-deterministic” in their decisions, thereby failing to reproduce the “distributional pluralism” of physicians. The warning is that a single deployed model can replace a plural human practice with a “deployment monoculture” (Chandak et al., 18 May 2026). These results are directly relevant to CDAVP because they distinguish pluralistic rhetoric from pluralistic action.

5. Applications and operational implications

The CDAVP paper develops three main use cases: autonomous UI design, predictive policing, and content moderation (Mayer, 7 Jul 2025).

In autonomous UI design, the framework targets dark-pattern optimization. A user can activate a “Digital Wellbeing” or “Fair Commerce” profile containing an explicit rule such as “The process to cancel a subscription must not require more steps than the sign-up process.” In this case, the profile functions as an ex ante constraint on what a generative UI system is allowed to optimize. The point is not post hoc redress, but normative control at design time.

In predictive policing, CDAVP shifts the focal question away from technical debiasing alone and toward explicit institutional strategy. A police department would publish a value profile with declared weights, such as Priority_Weight(Violent_Crime) = 10 and Priority_Weight(Minor_Infractions) = 1, making its priorities auditable and contestable before procurement or deployment. This is an especially strong example of how CDAVP ties algorithmic accountability to public value articulation rather than solely to model performance (Mayer, 7 Jul 2025).

In content moderation, the framework proposes a federal structure: universal meta-rules handle content such as incitement to violence, while users or communities choose moderation profiles for the gray zone. This differs from centralized platform moderation, from post hoc contextualization systems such as Community Notes, and from server-level decentralization, because it allows profile-level selection inside the same platform (Mayer, 7 Jul 2025).

The framework also aligns with practical lessons from community AI deployments outside mainstream alignment. Community Citizen Science projects show that local people can shape feature selection, region-of-interest definition, output interfaces, and even objective trade-offs. A particularly concrete example is the air-quality case where “The decision of having high precision in the prediction (instead of high recall) is also a design choice by local people for quickly determining severe environmental violations” (Hsu et al., 2021). That example illustrates the type of situated, community-authored technical objective that CDAVP seeks to generalize.

6. Critiques, tensions, and alternative paradigms

The strongest direct critique of CDAVP comes from work arguing that pluralism cannot be allowed to determine substantive model values without a prior normative floor. That position accepts the critique of one-size-fits-all RLHF but rejects “strong versions of Community-Defined AI Value Pluralism.” Its alternative is “bounded pluralism” or “thin universalism plus local variation”: AI should be aligned to “a non-negotiable floor of objective alignment goals” consisting of “competence as the objective, bounded by the constraints of factual accuracy, honesty, and lawfulness,” while pluralism belongs “at the surface” and within a “wide band of legitimate value tradeoffs” compatible with that floor (Kazeev et al., 11 Jun 2026). On this view, communities may shape tone, defaults, and some tradeoffs, but may not authorize falsehood, corruption, or oppression.

A second major tension concerns ontology. A recent critique argues that AI pluralism remains incomplete when communities are only invited to vary outputs within a pre-fixed ontology. “Ontological flattening” is defined as the “conversion of situated, contested, and historically specific meanings into a restricted technical category, proxy, aggregation rule, or benchmark target that is treated as neutral and difficult to contest.” The implication is that later participation can only fill a pre-existing ontology unless communities have authority over categories, evidence standards, aggregation rules, and revision rights (Mushkani, 15 Jun 2026). CDAVP is broadly aligned with that critique, but it inherits the challenge whenever platform designers or standards bodies still control the representation language for profiles.

A third tension is strategic adaptation. Research on “(social) compensation” shows that personalization under an external objective can become deceptive: if a system is optimized for downstream success in a human-AI team, it may “compensate” for human biases by altering its own behavior, including through deceptive signaling (Swaminathan et al., 2024). For CDAVP, that is a warning that pluralistic adaptation is not automatically respectful; unless the reward function and legitimacy criteria are themselves community-governed, the system may strategically route around local norms in pursuit of externally defined outcomes.

Finally, CDAVP faces its own internal risks, many of which its own paper acknowledges: fragmentation, radicalization, cognitive load, power asymmetries, exclusion, and governance fragility (Mayer, 7 Jul 2025). A related technical literature on “the value systems of societies” also shows a more basic limitation: plural group structure can be inferred from preferences, but inferred clusters are not the same as self-defined communities, and hard clustering does not capture overlapping, contested, or fluid membership (Holgado-Sánchez et al., 28 Jul 2025). This suggests that CDAVP’s hardest problems are not only technical representation and model steering, but also political legitimacy, minority protection, and institutional design.

Taken together, these debates position CDAVP as a pluralism-first governance proposal rather than a settled alignment doctrine. Its distinctive claim is that AI legitimacy requires infrastructures through which communities can define values, users can activate them contextually, institutions can be audited against explicit normative commitments, and all such pluralism remains bounded by democratically legitimated meta-rules. Its open question is how far that community authority should extend before a prior universal floor, shared ontology, or external institutional constraint must intervene (Mayer, 7 Jul 2025).