Community-Enriched AI: Models & Impact

• Community-Enriched AI is a paradigm that incorporates community knowledge, values, and participatory practices into AI systems, aligning technology with local norms and agency.
• It employs structured methodologies such as co-creation cycles, federated learning, and meta-rule governance—measured via indices like the DCLI—to enhance system trust and contextual relevance.
• The approach addresses challenges like power imbalances and tokenism by embedding continuous community oversight and mixed-initiative collaboration for transparent, adaptable AI deployment.

Community-Enriched AI refers to the systematic incorporation of community agency, knowledge, perspectives, and values into the design, development, deployment, and evaluation of artificial intelligence systems. Distinguished from top-down or developer-centric paradigms, Community-Enriched AI recognizes communities as co-designers, stewards, and validators of AI, seeking to align technical artefacts with locally situated norms, priorities, and lived experiences. This orientation sustains accountability, relevance, and equity by explicitly addressing issues of power, pluralism, and the sociotechnical factors that shape technology adoption and impact.

1. Conceptual Foundations and Definitional Scope

The notion of Community-Enriched AI encompasses several interrelated principles and models. In some contexts, it denotes participatory approaches where communities drive dataset creation, norm definition, or problem formulation—exemplified by the Wikibench system for Wikipedia, which empowers communities to curate, annotate, and govern AI evaluation datasets, surfacing consensus, dissent, and uncertainty explicitly within existing community workflows (Kuo et al., 2024). In AI education, Community-Enriched AI emerges through learning pathways shaped by learners’ identities, cultural practices, and histories of marginalization, leading to locally meaningful computational projects and the critique of automation-centric logics (Tena-Meza et al., 2021).

At a systems level, Community-Enriched AI integrates collective agency not just as a source of data or feedback but as an ongoing force in dataset curation, federated training (e.g., CommunityAI’s community-based federated learning architecture (Murturi et al., 2023)), algorithmic governance (Mayer, 7 Jul 2025), and civic engagement and oversight (Deng et al., 2023, Overney, 16 May 2025, Lin et al., 2024). In each, the community’s voice is foregrounded at multiple stages, requiring both technical and organizational infrastructures for deliberation, contestability, and ongoing adaptation.

2. Formal Models and Socio-Technical Architectures

Several formal models underpin Community-Enriched AI, varying by domain:

• Community-Defined Value Profiles: CDAVP (Mayer, 7 Jul 2025) encodes community values as structured, machine-readable profiles:
  • Each profile $p = (V_p, R_p, D_p)$ specifies weighted values, rights, and duties.
  • Users activate profiles contextually via $\alpha(u, s)$, triggering conflict detection and meta-rule-guided moderation.
  • Meta-rules, democratically ratified, mediate hard constraints and value conflicts across user- and community-activated sets.
• Federated and Distributed Learning: CommunityAI (Murturi et al., 2023) orchestrates federated learning by clustering clients into communities and cohorts based on metadata similarity, optimizing global and cohort-specific objectives:
  • Local objective: $$F_i(w) = \frac{1}{|D_i|} \sum_{x \in D_i} \ell(w; x)$$
  • Community-level aggregation with secure and privacy-preserving protocols, supporting heterogeneous devices and data schemas.
• Feedback Aggregation and Governance: In public engagement contexts, systems aggregate weighted community feedback through formulas such as

$$S = \frac{ \sum_{i=1}^n \sum_{j=1}^k r(u_i)\,w(e_j)\,f(u_i, e_j) }{ \sum_{i=1}^n \sum_{j=1}^k r(u_i)\,w(e_j) }$$

where $f(u_i, e_j)$ denotes feedback from user $u_i$ at event $e_j$, $w(e_j)$ weights event types, and $r(u_i)$ adjusts for demographic representativeness (Deng et al., 2023).

• Co-Liberation and Power Indices: In AI for Social Good partnerships, Lin et al. define the Data Co-Liberation Index (DCLI) to quantify the extent of co-leadership across project phases (Lin et al., 2024):

$$\text{DCLI} = \sum_{t=1}^5 w_t \cdot \alpha(t) \cdot p_c(t)$$

with $p_c(t)$ the community organization’s power share and $\alpha(t)$ the co-leadership intensity in phase $t$.

3. Methodologies and Participatory Workflows

Community-Enriched AI research and practice operationalize participation through well-defined methodologies:

• Co-Creation Cycles: Hsu et al. deploy multi-phase participatory cycles for community science, combining ethnographic immersion, participatory workshops, prototyping, iterative feedback, and negotiated model tuning (Hsu et al., 2021).
• PACT (Participatory Approach to enable Capabilities in communiTies): Bondi et al. merge the capabilities approach with participatory design across stakeholder identification, capability co-definition, co-design, embedded evaluation, and iterated adaptation, with equity and consensus metrics at each step (Bondi et al., 2021).
• Data Curation Platforms: Wikibench (Kuo et al., 2024) integrates plug-in labeling, consensus-building talk pages, and campaign management within native community workflows, enabling participatory dataset assembly with explicit disagreement and uncertainty tracking.
• Decentralized Organizing: Queer in AI exemplifies fully decentralized, intersectional participatory design, emphasizing flattened hierarchy, feedback loops, and reflexive critique to address power, access, and resource distribution (QueerInAI et al., 2023).
• Composite Narrative Synthesis: StoryBuilder involves iterative pipelines blending crowdsourced experience, LLM synthesis, and manual vetting to produce composite community narratives for civic engagement, with field and experimental validation of narrative framing impacts (Overney et al., 23 Sep 2025).

4. Design Patterns, System Features, and Evaluation

Community-Enriched AI systems exhibit recurring design patterns:

• Embedded Community Previews and Social Cues: Surfacing community-generated content and engagement metadata in LLM chatbot outputs—author attributions, vote counts, provenance—enables trust calibration and information foraging (Wang et al., 26 Jan 2026).
• Multi-Modal, Accessible Infrastructure: Deployment across physical touchpoints, digital dashboards, and in-person or asynchronous forums lowers barriers and supports participation diversity (“low floor/high ceiling”) (Deng et al., 2023, Saxena et al., 25 Feb 2025).
• Flexible Initiatives and Mixed-Initiative Collaboration: Interfaces support both human- and AI-driven contributions, with users retaining initiative over when and how AI is employed (Overney, 16 May 2025).
• Transparent Governance and Accountability: Versioned, auditable records of contribution, iteration, and decision provenance sustain legitimacy; meta-rules and participatory oversight formalize contestability and minimize power concentration (Mayer, 7 Jul 2025, Lin et al., 2024).
• Pluralistic Evaluation and Agency Metrics: Standard technical metrics (accuracy, precision, group-fairness), community-centric engagement metrics (participation rates, trust indices, feedback incorporation), and relational metrics (co-leadership index, agency ratio) are applied for multi-dimensional evaluation (Lin et al., 2024, Saxena et al., 25 Feb 2025).

5. Case Studies and Empirical Outcomes

Empirical studies validate the pragmatic impact and challenges of Community-Enriched AI:

• User Trust and Community Engagement: Community-enriched chatbots grounded in Kaggle posts (with social cues) increase both user trust and critical engagement compared to non-grounded LLMs, as measured by task performance, reliability ratings, and active exploration metrics (Wang et al., 26 Jan 2026).
• Participatory Value Alignment and Oversight: Implementations such as Barcelona Decidim and ACLU's CCOPS oversight ordinances demonstrate increased implementation rates for participatory budgets and delayed or redirected acquisition of surveillance technologies in response to community oversight (Saxena et al., 25 Feb 2025).
• Narrative Synthesis for Civic Dialogues: Experience-grounded composite narratives generated through a human-AI pipeline foster significantly greater respect and trust than theme-dominant or unsynthesized feedback, while multi-level interfaces and citations aid exploration and transparency (Overney et al., 23 Sep 2025).
• Participatory Dataset Curation: Community-driven labeling via Wikibench yields stronger alignment with community benchmarks on contentious classification tasks relative to developer-labeled counterparts, as well as rich representations of disagreement and low-confidence cases (Kuo et al., 2024).
• Power and Capacity Building in Social Impact Partnerships: Projects that exhibit high Data Co-Liberation Index scores (close to 1) entail greater community sustainability and utility, whereas low-scoring projects often fail to deliver usable outcomes and overburden community organizations (Lin et al., 2024).

6. Challenges, Limitations, and Future Directions

Several systemic and methodological barriers complicate the realization of Community-Enriched AI:

• Power Dynamics and Tokenism: Imbalances in participation, often entrenched by funding, institutional inertia, or technical gatekeeping, can lead to tokenistic engagement rather than genuine co-leadership (Hsu et al., 2021, Lin et al., 2024).
• Deliberation-Efficiency Trade-offs: Deep participatory methods introduce tensions between deliberative thoroughness and efficiency; hybrid approaches and computational mechanisms for surfacing high-disagreement cases are proposed (Kuo et al., 2024).
• Data Governance and Privacy: Scaling community-driven federated learning and participatory data curation raises non-trivial privacy, security, and representation challenges, requiring strict protocols, auditing, and oversight (Murturi et al., 2023).
• Sustainability and Long-Term Maintenance: Sustained funding, community capacity building, and infrastructural support remain uncommon, risking project abandonment or loss of community agency over time (Hsu et al., 2021, Lin et al., 2024).
• Pluralism and Contestability at Scale: Infrastructuring for explicit value pluralism and contestability—across overlapping community profiles, machine-readable value sets, and dynamic meta-rule layers—remains a frontier for HCI and AI governance research (Mayer, 7 Jul 2025).
• Evaluation and Accountability: Robust frameworks for measuring not solely technical merit, but also the legitimacy, equity, and relational impacts of community-enriched interventions are urgently needed (Saxena et al., 25 Feb 2025, Lin et al., 2024).

Future research directions include hybrid crowdsourcing–AI pipelines for capturing pluralistic data, formal audits for power and labor visibility, development of domain-specific meta-rule protocols for value conflict moderation, and participatory toolkits integrating deliberative, technical, and governance activities across lifecycles and scales.

7. Summary Table: Dimensions of Community-Enriched AI

Dimension	Example Implementation/Metric	Source
Co-Design/Co-Leadership	DCLI, CLI, participatory narratives	(Lin et al., 2024, Overney et al., 23 Sep 2025)
Dataset Curation & Evaluation	Disagreement σ, IAA κ, transparency	(Kuo et al., 2024)
Value Pluralism & Meta-Rules	Profile registry, meta-rule voting	(Mayer, 7 Jul 2025)
Federated Community Modeling	Cohort-based FL, privacy, adaptation	(Murturi et al., 2023)
Participatory Narratives	Composite stories, human+AI curation	(Overney et al., 23 Sep 2025)
Power/Agency Metrics	Co-leadership, burden equity, capacity transfer	(Lin et al., 2024)
System Sustainability	Maintenance planning, community training	(Hsu et al., 2021, Lin et al., 2024)

Community-Enriched AI marks a shift from algorithm-centered optimization toward democratic, pluralistic, and agency-driven approaches, embedding communities as persistent architects and governors of AI systems. This reorientation, grounded in formal models, participatory workflows, and empirical impacts, provides both practical toolkits and critical frameworks for the equitable integration of AI into social, civic, and scientific domains.