Participatory AI: Practices & Governance

• Participatory AI is a field that involves communities in the design, development, and governance of AI systems to enhance fairness and accountability.
• It employs diverse methods—from co-design and crowdsourcing to modular system integration—to incorporate local knowledge and reshape decision-making.
• It transforms the AI lifecycle with robust governance models and participatory recruitment strategies that rebalance power and ensure transparent oversight.

Participatory AI denotes AI design, development, deployment, and governance practices that meaningfully involve communities and publics beyond technical designers, especially those affected by the system. Across the recent literature, it is treated not as a single method but as a broad family of practices spanning problem formulation, data collection, model development, evaluation, documentation, deployment, and post-deployment oversight. Its central rationale is simultaneously normative and practical: participation is presented as a way to rebalance decision-making authority, enhance fairness and legitimacy, incorporate lived experience and local knowledge, and improve the quality, accountability, and real-world performance of AI systems (Birhane et al., 2022, Parthasarathy et al., 2024).

1. Genealogy, political character, and democratic commitments

Participatory AI is situated within a longer genealogy of participation in democracy, international development, healthcare, environmental decision-making, social movements, workplace democracy, and socio-technical design. A recurrent historical reference is Scandinavian participatory design from the 1970s–80s, especially the Collective Resource Approach and the Socio-Technical Systems Approach, which linked participation to democracy, autonomy, and collective management rather than to consultation alone (Birhane et al., 2022). A later Scandinavian formulation explicitly treats AI as “as much a political artifact as a technological one” and rearticulates Participatory AI through five participatory design principles—Mutual Learning, Future Alternatives, Artifact Ecologies, Empowerment and Mediation, and Emancipatory Practices and Democracy—together with four design challenges for algorithmic automation: Specialization, Multimodality and Distributed Ecologies, Emergent System Behavior, and Human Augmentation (Elmqvist et al., 16 Sep 2025).

This political framing is central. The literature repeatedly argues that AI systems are not neutral tools: they shape who decides what counts as valid knowledge, whose data and labor are used, whose values are optimized, and who benefits from deployment. For that reason, participatory AI is commonly presented as a democratic response to opaque, centralized, expert-led, and overly generic AI systems, particularly where they influence work, education, civic life, healthcare, public services, and everyday social relations (Elmqvist et al., 16 Sep 2025).

At the same time, participation is treated as historically ambivalent rather than intrinsically emancipatory. One widely cited warning is that “participation without redistribution of power is an empty and frustrating process for the powerless.” In this line of work, participatory practices can be emancipatory, but they can also legitimate already unequal institutional arrangements, mask extraction, or function as a veneer for control (Birhane et al., 2022). A related abstract on “Expansive Participatory AI” argues that participatory AI can contribute to democratizing the design of technology by shifting attention toward what should be designed, yet existing institutional power dynamics hinder the realization of expansive dreams and aspirations of relevant stakeholders, particularly in work with youth (Chang et al., 2022).

2. Conceptual boundaries and forms of participation

A major theme in the literature is that participation should not be reduced to a vague label. Several papers distinguish participation from inclusion, consultation, data labor, documentation work, and democratic governance. Inclusion means being present or accommodated; participation means actively shaping outcomes. Consultation can solicit input without redistributing authority. Data labor can improve models without benefiting contributors. Participatory AI is also not treated as a substitute for democracy or regulation, especially where systems are coercive or operate in the public interest (Birhane et al., 2022).

One influential typology organizes participatory AI into three overlapping categories. Participation for algorithmic performance improvement uses annotation, cleaning, labeling, feedback, and microtasks to improve datasets, model accuracy, robustness, diversity, or data quality. Participation for process improvement includes participatory design, user-centered design, co-design, human-in-the-loop model development, citizen science, crowdsourcing, surveys, focus groups, and UX studies; this mode can incorporate lived experience but often leaves overall goals fixed by designers or institutions. Participation for collective exploration is the most expansive form: communities help define the problem, objectives emerge collectively, co-learning and deliberation are central, refusal is possible, and the project can be reshaped or abandoned (Birhane et al., 2022).

Subsequent work has sharpened this critique by arguing that participation often begins too late. “AI From the Margins” states that participation typically starts after problem definitions and success criteria have been set, leaving limited room for minoritized communities to reshape what an AI system is for. It therefore introduces a preparatory methodological stance with seven preconditions: Reciprocity, A decolonizing stance, Centering minoritized standpoints, Methodological flexibility, Structural availability of refusal, Multidimensional accessibility, and Co-ownership. In this formulation, lived experience is not one input among many but an epistemic standpoint from which harms become legible (Portegies et al., 31 May 2026).

A parallel extension concerns stakeholders who are not direct end users. Work on secondary stakeholders in AI defines meaningful participation through three ideals—informedness, consent, and agency—arranged as a stepwise ladder. This account broadens participatory AI beyond primary stakeholders such as users or research subjects to include moderators, data contributors, activists, practitioners, and others who influence or are influenced by AI without a direct contractual relationship to the system (Ajmani et al., 8 Jun 2025).

3. Lifecycle architectures and governance models

Participatory AI is increasingly articulated as a lifecycle and governance problem rather than only a design method. One principled governance model divides the AI lifecycle into Design, Development, Deployment and Adoption, and Post-deployment monitoring / feedback loops. Participation is mapped differently across these stages: non-experts and affected communities are especially valuable in deciding whether AI is needed at all, shaping the problem statement, understanding social context, and identifying harms early; technical experts dominate model development; and deployment requires risk assessment, governance and legal safeguards, grievance channels, monitoring, and feedback loops from users and affected persons (Parthasarathy et al., 2024).

The same work proposes a “decision sieve,” a layered model through which decisions and information pass at each lifecycle stage. Stakeholder identification is organized through four criteria—Power, Legitimacy, Urgency, and Harm—while stage assignment prioritizes Urgency first, then Expertise, and Power last. It also distinguishes horizontal transfer of information within a stage from vertical transfer across stages, arguing that transparency and communicability are central to the former, while interpretability and explainability matter for the latter. After participatory inputs are collected, responses are to be collated stage-wise either through Voting / aggregation or Sorting / consensus building (Parthasarathy et al., 2024).

A complementary proposal re-architects the conventional engineering pipeline into an “augmented AI lifecycle” with five interconnected phases: co-framing, co-design, co-implementation, co-deployment, and co-maintenance. Here participation is continuous rather than episodic. Co-framing jointly defines the problem and surfaces contextual risks before technical work begins; co-design negotiates data sources, model families, interfaces, and trade-offs; co-implementation opens training and validation artifacts to participatory review; co-deployment adds governance charters, dashboards, and recourse; and co-maintenance institutionalizes periodic technical, ethical, and participatory audits, including the possibility of suspension or decommissioning (Mushkani et al., 31 Jul 2025).

Public-sector work extends this logic into procurement and implementation. A workshop agenda on public sector innovation defines participatory AI as public participation and community engagement in the scoping, design, adoption, and implementation of public sector algorithms. It argues that higher standards apply in government because AI is used for urban planning, security, surveillance, energy and critical infrastructure, and other functions that directly affect citizens’ access to essential services. In that context, procurement and vendor contracts become a crucial site of participatory governance, since agencies often acquire rather than build AI systems (Saxena et al., 25 Feb 2025).

4. Power, recruitment, refusal, and representational accountability

Across the literature, meaningful participation is consistently linked to redistribution of power, reciprocity, reflexivity, community benefit, the ability to refuse or reshape the project, and long-term contextual engagement. Participation is treated as hollow when it solicits input without affecting decision-making, when benefits accrue mainly to institutions or corporations, or when dissent and refusal are not structurally available. This is why the literature repeatedly warns about co-optation, participation-washing, confusion between participation and labor extraction, and the use of participatory language to legitimate surveillance or harmful deployment (Birhane et al., 2022).

Recruitment is therefore not described as a neutral logistical pre-step. A study of 37 participatory AI projects and interviews with 5 AI researchers defines recruitment methodology as the process of identifying, reaching out to, and engaging stakeholders. It codes projects by who is recruited, who recruits or initiates, recruitment strategy, participant empowerment, and documentation completeness. Four strategy types are identified—Organizational recruitment, Infrastructural recruitment, Personal networks, and Events—with organizational recruitment the most common pattern. The same study argues for “relationship-forward recruitment” and “reflexive recruitment documentation,” emphasizing that recruitment shapes who gets a seat at the table, how trust is built, and whether participation becomes tokenistic or empowering (Kim et al., 27 Aug 2025).

Refusal is not only a normative principle but an operational one. The Māori data-rights case in participatory AI presents refusal and restricted sharing as legitimate participatory acts, tied to data sovereignty, community control, and the insistence that downstream technologies must directly benefit Māori people and be built by Māori (Birhane et al., 2022). “AI From the Margins” likewise treats refusal of AI itself as a legitimate methodological outcome rather than a failure of participation (Portegies et al., 31 May 2026). Work on co-constructed alignment similarly includes non-engagement, withdrawal, and limiting use as valid alignment strategies rather than as defects in user behavior (Arzberger et al., 22 Jan 2026).

Recent work on AI-mediated public consultation has added a measurement framework for whether public input survives institutional summarization. “Participatory provenance” shifts audit from output quality to the “chain of custody” from individual submission to official summary. Applied to Canada’s 2025–2026 national AI Strategy consultation, it reports that both official government summaries underperform a random-participant baseline, with $-9.1\%$ and $-8.0\%$ coverage degradation, and that 16.9\% and 15.3\% of participants were effectively excluded. Exclusion concentrated in clusters expressing dissent, scepticism, and critique of AI, with exclusion rates from 33–88\% (Mahajan, 22 Apr 2026). This work makes representational fidelity a central participatory criterion: the question is not only who was invited to speak, but whose voice remained legible in the institutional record.

5. Community-led practice and sectoral instantiations

Participatory AI has been instantiated in community-led organizations, educational design, organizational decision-making, journalism, and civic participation. A prominent community-led case is Queer in AI, presented as a grassroots, volunteer-run, global, decentralized organization organized around decentralized organizing, intersectionality, and community-led initiatives. It is described as building aid and programming “by and for” the queer community, while also influencing conferences, publishers, companies, and institutions. The organization is reported to have grown to around 870 members across more than 47 countries, and its Graduate School Application Fee Aid Program provided 31 recipients, $16,689 aid** in 2020/2021, **81 recipients, $70,607 aid in 2021/2022, and 48 recipients, $40,476 aid in 2022/2023 at the time of writing (QueerInAI et al., 2023).

Educational work extends participation to children and teachers. A challenge-based participatory design study at IDC 2023 collected 60 design ideas from children aged 4 to 16 (mean age 9.80, SD 1.76) on the theme “Smart Communities: Rebuilding a compassionate world!” The ideas focused on helping people, protecting the environment, engendering kindness, supporting connections, and facilitating equality, with intelligent agent concepts comprising 71\% of applications. The study treats children as co-designers and informants rather than passive recipients of AI, while also stressing the dependence of child participation on adult facilitation and AI literacy scaffolding (Bai et al., 2023). A separate education case in the Scandinavian participatory AI literature describes teachers, educational researchers, NLP researchers, and digital humanities researchers co-designing school-subject-specific AI activities through workshops, prototyping, co-teaching, and ongoing communication, with emphasis on teacher agency and students’ ability to contest AI output (Elmqvist et al., 16 Sep 2025).

In organizational decision-making, “Deliberating with AI” uses ML models as boundary objects for reflecting on admissions processes rather than automating them. In a graduate admissions case study, 7 faculty participants and 9 student participants worked with an anonymized historical dataset of 2207 applicants from Fall 2013 to Fall 2019. Participants explored data, selected features, trained models, and evaluated performance and fairness. The study reports that students generally preferred models with more false positives, while faculty preferred fewer false positives, making divergent fairness intuitions and organizational values explicit (Zhang et al., 2023).

Workplace and newsroom settings raise related questions of agency. In journalism, interviews with 10 participants in an independent local newsroom led to the Gradual Voluntary Participation (GVP) framework, which defines participation as gradual, voluntary, and organizationally embedded. Its five principles are Recognition Without Coercion, Continuous Value Alignment, Contextual Knowledge Scaffolding, Sovereignty Preservation, and Transparent Accountability, and its matrix maps participation along depth and scope rather than a single ladder (Barbini et al., 23 Apr 2026). This responds to what the paper calls a “perception gap,” in which trust in AI depends less on nominal inclusion than on whether journalists experience real agency in workplace participatory workflows.

Civic and public-sector cases broaden the domain further. A São Paulo case study of participatory budgeting describes an eight-phase hybrid process and reports that public voting fell from 43,000 votes in 2021 to 6,070 votes in 2025, an 86\% decline, with participation amounting to only 0.05\% of the city’s population in 2025. The study analyzes how AI could support analysis and systematization of demands, facilitation of dialogue and deliberation, inclusion and accessibility, and visualization and simulation, while insisting that AI should support, not replace, human participation (Sousa et al., 20 Sep 2025).

6. Technical systems, formalization, and emerging directions

An important recent shift is the translation of participation from a social principle into a technical systems paradigm. “Scaling participation” proposes modular AI systems in which diverse stakeholders contribute small LLMs that remain separate components in a collaborative system. The study contacted 236 researchers and obtained 61 models, then formed collaborative systems from 2, 4, 8, 16, and 32 models. It reports that participatory modular systems outperform non-modular and/or non-participatory baselines by up to 15.4\% across 15 tasks, that scaling from 2 to 32 models yields an average improvement of 28.78\%, and that collaborative systems solve over 15\% of problems where all individual models fail (Feng et al., 5 Jun 2026).

This work also introduces compositional strength as a distinct evaluative concept: a model’s value lies not only in its individual performance but in the performance gain it contributes when added to a collaborative system. The paper reports that individual performance does not correlate strongly with compositional strength, with Pearson correlation not significant and p-values consistently larger than 0.05 across three collaboration algorithms and two datasets (Feng et al., 5 Jun 2026). This is a substantial departure from conventional model-centric evaluation, and it suggests a technical research agenda in which participatory contribution is measured at the level of system composition rather than standalone capability.

Other technical systems make participation legible through interfaces and orchestration. A participatory differential privacy framework for public decision-making uses a conversational interface with an adaptive $\epsilon$-selection protocol based on TOPSIS, real-time Mean Absolute Error (MAE) visualizations, GPT-4-powered impact analysis, and a legal-compliance mechanism that constrains privacy budgets within $[0.1, 2.0]$. In simulation on the Household Electricity Demand dataset, it reports a strong negative correlation between $\epsilon$ and MAE, $r = -0.96$, $p < 0.01$, and states that privacy-first configurations introduce 3.6× more noise than utility-optimized settings (Yang et al., 30 Apr 2025). Here participation is framed as negotiation scaffolding: stakeholders can influence mathematically consequential privacy settings without violating legal and technical bounds.

Urban and collaborative design systems add multi-user and multi-agent structures. CoDesignAI is a web-based system for conceptual-stage urban design that combines multiple users with multiple AI agents, including an AI facilitator and optional AI expert agents such as an AI urban designer and an AI urban planner. The system integrates conversational AI with Google Maps Platform and Google Street View, uses a round-based workflow, and converts discussion into concise implementation-oriented prompts for image revision grounded in a real street scene (Zhang et al., 16 Mar 2026). The paper presents this as a proof of concept for shifting urban design from an expert-centered practice toward a more open and participatory process.

Finally, participatory AI is being extended into runtime interaction and alignment. “Co-Constructing Alignment” defines co-construction as treating users as active contributors to alignment at run-time, drawing on situated sense-making of local misalignments rather than receiving pre-specified values. Using misalignment diaries and generative design activities with researchers using LLMs as research assistants, it finds that misalignments are experienced less as abstract ethical violations than as unexpected responses, and task or social breakdowns; participants articulate roles that include adjusting, grounding, questioning, reflecting, limiting use, and deliberate non-engagement (Arzberger et al., 22 Jan 2026). This work reframes alignment as an ongoing, situated, and shared practice rather than a property fixed entirely upstream.

Participatory AI therefore encompasses a heterogeneous but increasingly coherent research field: democratic and community-led design, power-sensitive governance architectures, recruitment and accountability methods, participatory public consultation, and technical systems in which participation shapes models, interfaces, privacy parameters, collaborative design workflows, and runtime alignment. A common thread is that participation is treated as meaningful only when it changes who can define problems, which information counts, how trade-offs are negotiated, and who retains the authority to contest, refuse, or reconfigure AI systems.