AI for citizen participation in international policymaking: a collaborative governance perspective on multilingual inclusion

1 Introduction

A traditional top-down approach to policy design, where institutional actors and experts predominantly set an agenda and then search for a solution, is increasingly seen as technocratic and out of touch with citizens’ lived experiences (Brinkman et al., 2023). This democratic deficit is particularly acute in global governance, where recent UNFCCC COP processes and WHO pandemic treaty negotiations revealed structural barriers to meaningful civic engagement, including underrepresentation of stakeholders from developing countries and a reliance on informal, English-only meetings that marginalized non-Anglophone speakers (Schwalbe et al., 2025; Sultana, 2022). The challenges are compounded by the scale, linguistic diversity, and institutional distance inherent to international policymaking–precisely the conditions that make sustained, reciprocal engagement between citizens and public officials most difficult to achieve (Hilbrich, 2022; Dahl, 1999).

Advances in artificial intelligence offer promising opportunities for addressing these barriers. AI tools can substantially lower the cost of large-scale consultations, enabling automated translation, topic modelling, and sentiment analysis of thousands of citizen inputs that manual methods cannot match (OECD, 2023; Shin et al., 2024; Ustenko et al., 2026). Recent applications include AI-assisted participatory platforms that cluster citizen proposals and recommend policy options (Arana-Catania et al., 2021), legislative summary tools that facilitate citizen commentary on draft laws, and deliberative platforms that map agreement and disagreement across large participant pools in real time (Landemore, 2023). Yet evidence on effectiveness remains mixed. E-participation research identifies a persistent gap between technological promise and democratic outcome: platforms that reach many citizens do not reliably generate the quality, depth, or representativeness of input that legitimate governance requires (Toots, 2019).

Two literatures address this gap, but have been developed largely in parallel. First, collaborative governance research has identified the institutional design features that enable effective citizen engagement in policy-making, centred on the conditions that generate and sustain shared motivation among participants (Ansell and Gash, 2007; Emerson et al., 2012). Second, an emerging AI-for-deliberative-democracy literature has emphasised the technical features that support democratic values, including equity, inclusivity, and transparency (Landemore, 2023; Shin et al., 2024; Williams, 2026). The integration of these two bodies of knowledge, specifying when and how AI tools support the foundations of collaborative governance, remains underdeveloped. This gap is especially consequential in the international policy-making context, where (1) the non-electoral, non-binding character of consultation processes removes the disciplining mechanisms that anchor policymaker attention to citizen concerns in national settings, and (2) where the institutional distance between citizen input and policy output makes the risks of symbolic, rather than substantive inclusion particularly acute (Hilbrich, 2022; Fung, 2006).

To address this gap, this paper proposes an analytical framework for assessing how AI-assisted participation operates across the collaborative governance process (Figure 1). The framework is grounded in Emerson et al.’s (2012) integrative model for collaborative governance and conceptualizes principled engagement as a staged process comprising initiation (bringing relevant actors into the process and lowering barriers to entry), discovery (identifying and organizing the concerns, experiences, and priorities participants bring), deliberation (enabling participants to exchange views, compare perspectives, and weigh alternatives), and determination (translating citizen input into policy outputs or recommendations). Through these stages, shared motivation–trust, mutual understanding, internal legitimacy, and commitment among participants–is formed and sustained. Within this framework, AI is conceptualized as a stage-specific enabling condition that can strengthen or weaken shared motivation. For instance, AI-assisted translation at initiation can lower language barriers and broaden access therefore strengthening trust in the participatory process, but if the same tools are trained predominantly on Global North corpora, they may misread Global South responses as less coherent, reproducing the exclusions they were designed to address (Joshi et al., 2020).

The framework is illustrated empirically through the Civil 20 (C20) India 2023 International Grassroots Survey–a large-scale citizen consultation conducted as part of the C20 policy deliberation process, involving 5,152 participants from 64 countries and designed to include underrepresented groups. The survey was conducted in 24 different languages. The empirical illustration focuses on the initiation and discovery stages; the deliberation and determination stages are addressed as theoretical propositions. At the initiation stage, subgroup analysis of participation quality reveals whether the consultation’s reach translated into substantive voice or whether systematic inequalities across demographic groups left some participants formally present but effectively unheard. At the discovery stage, natural language processing algorithms assess whether shared topics carry shared meanings across communities, or whether the diversity of frames through which different groups articulate common concerns is suppressed by aggregate topic prevalence metrics. Frames here are understood as interpretive structures that organise how a problem is understood and what kinds of responses appear appropriate (Entman, 1993).

Together, the analytical framework and empirical illustration address two research questions:

RQ1: What conditions in the design and deployment of AI tools determine whether AI-assisted citizen participation strengthens or weakens shared motivation at each stage of the collaborative governance process?

RQ2: What patterns of participation inequality and frame diversity pose risks to shared motivation in the multilingual context of the C20 India public consultation?

The paper proceeds as follows. Section 2 develops the analytical framework, introducing Emerson et al.’s collaborative governance model, specifying AI as a stage-specific enabling condition within principled engagement, and deriving four stage-specific propositions. Section 3 presents the empirical illustration focused on the initiation and discovery stages of the C20 public consultation. Section 4 discusses the findings in relation to the framework and their implications for the design of AI-assisted global governance consultations.

2 Analytical framework

2.1 AI for citizen participation

Contemporary scholarship on democratic innovations increasingly views AI as a potential solution to the “deliberative trilemma” of mass participation, political equality, and meaningful deliberation (Landemore, 2023; Tessler et al., 2025). Yet systematic analyses of existing digital participation tools revealed several persistent limitations (Kanagaraj et al., 2025). First, these technologies create a deficit in accountability: while effective at channeling citizen input toward governments, they rarely communicate back to participants how their contributions factored in policy decisions (Shin et al., 2024). Second, a narrow focus on technical solutions for scaling participation (e.g., topic modelling) may neglect the nuanced requirements of high-quality deliberation and risk reinforcing systemic biases if not explicitly designed to empower marginalised groups (Shortall et al., 2022; Taylor et al., 2024). This is especially important in multilingual settings, where AI tools can widen access by translating contributions across languages, but may also flatten culturally specific meanings (Naveen and Trojovský, 2024).

These practical limitations partly stem from deficiencies in theoretical foundations. Existing research most commonly evaluates AI-assisted participation at the tool level, without disaggregating effects across the sequential stages of a governance process or sufficiently addressing institutional drivers of meaningful policy participation. Taking a collaborative governance perspective helps to address this gap since its focus is precisely on the relational and institutional conditions that make participation meaningful rather than merely procedural.

2.2 Collaborative governance and shared motivation

Collaborative governance is a structured process through which public agencies, citizens and other non-state stakeholders work together to make decisions and pursue public goals that cannot be achieved by one actor alone (Ansell and Gash, 2007; Emerson et al., 2012). Emerson et al. (2012) identify several elements of collaborative dynamics. The first element is principled engagement, which stands for the iterative process through which participants discover shared interests, define common problems, deliberate over alternatives, and reach determinations. The second element is shared motivation, which refers to the interpersonal and relational foundations of collaboration and can be broadly understood as a form of social capital (Coleman, 1988; Putnam, 2000). Shared motivation comprises four interdependent dimensions: trust in the process and other participants; mutual understanding of differing perspectives; internal legitimacy, or participants’ sense that the process is credible and their involvement meaningful; and commitment to acting on collectively reached conclusions.

In Emerson et al.’s original model, shared motivation is developed and sustained mainly through direct human interaction: participants build trust by working together, develop mutual understanding through dialogue, earn internal legitimacy through inclusive process design, and commit through shared investment in outcomes. The quality of principled engagement directly determines the strength of shared motivation, and the two reinforce each other in a virtuous cycle.

However, citizen engagement in international policymaking typically relies on remote and episodic participation–conditions which are less conducive to development of shared motivation (Ansell and Gash, 2007; Fung, 2006). A further consideration is that forums like the G20 and C20 operate outside the disciplining pressure of electoral accountability, which usually effectively brings policymakers’ attention to the immediate concerns of the citizens (Macdonald and Macdonald, 2006; Slaughter, 2013). Multilingualism adds another layer of difficulty, because linguistic diversity can make mutual comprehension harder to achieve due to the reliance on translation (Cabrera, 2024; Verhasselt, 2025). The present framework addresses these compounding challenges by specifying, stage by stage, the conditions under which AI strengthens rather than undermines the shared motivation for collaborative governance.

Emerson et al. (2012) specify four elements of principled engagement: discovery, definition, deliberation, and determination. The framework developed here retains discovery, deliberation, and determination, adds initiation as the first stage of principled engagement, and treats definition as distributed across the process rather than located in a single stage. The initiation stage elevates considerations Emerson et al. (2012, p. 11) address in their paper: who participants are, how they are selected, and how many there are. However, their model treats participant entry as a background consideration, at least partly because in conventional collaborative governance settings engagement is ongoing and open to admitting different stakeholders at different points. A one-shot international consultation normally has no such continuation, so composition and the depth at which participants can speak are settled once and carried forward almost unrevised. This is reinforced by Emerson et al.’s own rationale for inclusion, since they hold that diversity gives voice to multiple perspectives and that participants’ relative power can enable or disable subsequent agreements.

In Emerson et al.’s original formulation, definition consists of settling what participants are collectively trying to achieve, what vocabulary they will use, what they expect of one another, and what standards they will apply to information. Each of these assumes sustained interaction between participants, and citizen engagement in international policymaking, as noted above, typically limits that interaction. The definitional work does not disappear under these conditions; it is performed on participants’ behalf and at different points in the process. For instance, a shared vocabulary is set through translation and topic modelling, and criteria of relevance are set by whichever body determines which citizen frames are carried forward into policy outputs. Therefore, treating the definition not as a standalone stage but as a function distributed across the process shifts the analytical question. Rather than asking whether a definitional stage was well executed, the framework asks who exercised definitional authority, at what point, and with what effect. Those answers help specify the design conditions AI-assisted tools must meet.

Several elements of Emerson et al.’s original model are treated here as given rather than analysed: the system context, and the drivers of collaboration, namely leadership, consequential incentives, interdependence, and uncertainty. These bear on principled engagement but are not themselves conditions AI can act upon directly, and holding them fixed keeps the focus on the conditions of AI use that arise within principled engagement itself during a multilingual international consultation. Leadership, for instance, may shape who enters a consultation and what levels of trust and perceived legitimacy they bring to it, which raises the further question of how far AI-assisted facilitation might rebuild trust where leadership is weak. That question is left open here. Capacity for joint action, the third component of collaborative dynamics alongside principled engagement and shared motivation, is likewise outside the present scope. Emerson et al. (2012, p. 14) describe it as “the link between strategy and performance,” comprising the procedural and institutional arrangements, leadership, knowledge, and resources needed to sustain an ongoing collaborative regime. The international consultations examined here typically do not produce such a regime.

The four stages developed below, initiation, discovery, deliberation, and determination, specify the conditions under which AI-assisted tools strengthen or weaken shared motivation within this scope.

2.3 Stages of AI-assisted principled engagement

2.3.1 Initiation

The initiation stage is where citizens enter the consultation process and where the conditions for principled engagement are first established (Figure 1). The primary function of this stage is ensuring sufficiently broad and diverse participation (Carlson, 2007). The shared motivation dimensions most at stake here are trust and internal legitimacy. While all four dimensions of shared motivation operate throughout the collaborative governance process, trust and internal legitimacy function as preconditions for the others: without an initial sense that the process is credible and accessible, the mutual understanding and commitment that develop through later stages cannot take hold (Ansell and Gash, 2007; Fung, 2006).

The current research provides substantial evidence corroborating that AI tools may strengthen trust and internal legitimacy of participatory processes by lowering the physical and logistical barriers that have historically excluded groups defined by linguistic, geographic, or demographic constraints (Zuiderwijk et al., 2021). Neural machine translation enables multilingual consultation at a scale and cost that human translation cannot match, as illustrated by Lohar et al. (2023) deliberative system, which enabled participation across six European languages. AI-assisted platform design can further lower participation thresholds by adapting interfaces to varying levels of digital literacy, which was cited as a major obstacle to civic engagement (Li and Li, 2022; Van Dijk, 2006). Beyond physical access, conversational agents probing for missing details and generating structured summaries of contributions can support low-literacy and non-expert participants in articulating their concerns better (Tavanapour et al., 2019; Sheshadri et al., 2025).

Whether these tools genuinely strengthen shared motivation, however, depends on a distinction political science has long recognised. Young (2000) distinguishes external exclusion, where citizens are kept out entirely, from internal exclusion, where citizens are nominally present but effectively silenced because deliberative norms privilege “dispassionate, orderly, or articulate” styles of expression (Young, 2000, p. 6). In large-scale public consultations, this distinction requires attending to both the demographic composition of participants and the quality of their contributions - specifically, whether responses are sufficiently complete and detailed to be legible to downstream processing. When participation counts by subgroups are reported as the only indicator of process inclusivity, citizens with lower response completeness may be formally included but substantively underweighted. This form of exclusion from the policy process has been shown to erode trust (Nabatchi, 2010; Schmidthuber et al., 2021).

These concerns about inclusion are not unique to AI-assisted contexts. Lima and Podadera (2026), developing the Citizen Participation Index across six countries, document that internet access, education, literacy, and linguistic diversity operate as significant constraints on participation potential, with substantial disparities driven by socio-economic and cultural variables. AI-assisted initiation tools therefore enter a context where these constraints already shape who can participate substantively, and design choices must contend with them rather than assume they have been resolved.

These considerations together motivate the following proposition:

P1: Variation in participation quality across demographic groups creates a baseline inequality that aggregate inclusion metrics alone cannot detect and that AI-assisted initiation tools must explicitly address to avoid undermining trust and internal legitimacy.

2.3.2 Discovery

The discovery stage is where citizen inputs are processed, organised, and made usable to analysts and policymakers. Its main objective is to produce a representation of citizen frames that reflects the range and depth of what participants expressed. The key shared motivation dimension that should emerge at this stage is mutual understanding. When discovery produces a flattened or biased representation, the basis for genuine mutual understanding may be weakened (Dryzek, 2000).

AI-assisted text analysis can strengthen mutual understanding at the discovery stage by surfacing patterns across thousands of responses that would be hard for human analysts to identify at comparable scale. Topic modelling methods, including Latent Dirichlet Allocation (LDA) (Blei et al., 2003) and transformer-based embeddings (Raman et al., 2025; Gusha et al., 2025), can identify thematic clusters within large datasets and help organise open-ended citizen input for subsequent analysis. Keyword-based and sentiment analysis can further enrich this picture by showing how similar topics are framed differently across groups, offering a more textured account of citizen experience than topic prevalence alone (Törnberg and Törnberg, 2016; Gabrielatos, 2018).

Whether these tools strengthen mutual understanding, however, depends on a distinction between what citizens are talking about and how they articulate it. Grimmer and Stewart (2013) show that computational text analysis does not simply recover structure from political text; it imposes analytical choices that can privilege tractability over meaning. Roberts et al. (2014) further suggest that standard topic models may obscure differences between how much participants discuss a topic and the register through which they do so, which can limit their ability to capture how different groups express shared concerns differently. In deliberative terms, this matters because mutual understanding and legitimacy depend not only on identifying dominant themes, but on preserving the plurality of discourses through which political concerns are expressed (Dryzek, 2000).

A particular challenge at the discovery stage in multilingual consultations is the handling of conflicting frames that emerge across language communities. Translation may render lexically equivalent terms across languages, but the conceptual frames those terms carry can diverge substantially: what registers as a concern about “safety” in one linguistic and cultural context may map onto crime, food security, or labour precarity in another. Aggregating across languages risks merging distinct concerns into single thematic categories that obscure these differences. The discovery stage must therefore handle conflicting multilingual frames not by harmonising them prematurely, but by preserving them as distinct inputs into deliberation, allowing participants and policymakers to engage with the linguistic and cultural specificity of how concerns are expressed (Verhasselt, 2025; Cabrera, 2024).

In large-scale public consultation, discovery must therefore attend not only to topic prevalence, but also to subgroup variation in how shared concerns are articulated. When aggregate topic prevalence is treated as the primary summary metric, groups with different lived experiences may appear to prioritise the same issue while expressing it through systematically different conceptual and emotional registers. Recent reviews suggest that computational text analysis in public participation is used primarily to identify, classify, and summarize themes in citizen input, whereas systematic attention to subgroup-level heterogeneity remains less developed (Romberg and Escher, 2024). Based on these considerations, the following proposition is formulated:

P2: AI-assisted discovery of citizen perspectives strengthens mutual understanding only when it moves beyond aggregate topic prevalence to subgroup-level frame analysis.

2.3.3 Deliberation

The deliberation stage is where participants exchange reasons, challenge each other’s perspectives, and work toward shared understanding of the issues. This stage is supposed to ensure that the full range of views, including dissenting and minority positions, has been taken into consideration and engaged with. This form of deliberation can further develop mutual understanding and it has traditionally been possible only in small groups.

AI tools are increasingly proposed as a means of extending deliberation to mass participation (Landemore, 2023). Klein (2012) demonstrates this with the Deliberatorium, showing that attention-mediation metrics can direct participants toward high-priority areas of a structured deliberation map. Platforms such as Polis similarly use computational clustering to map areas of agreement and disagreement across large participant pools (Small et al., 2023). The vTaiwan process further demonstrates how algorithmically mediated public deliberation can operate at the level of national policy, combining online and offline consultation to inform digital legislation in Taiwan (Hsiao et al., 2018). In multilingual settings, AI-based translation may further extend deliberative inclusion by enabling cross-linguistic exchange, a capability Landemore (2023) identifies as one of AI’s most significant contributions to diverse and inclusive participation.

However, the application of AI for facilitating deliberation poses certain risks. Recent work on LLM-assisted deliberation shows that summarisation can enable new forms of collective sense-making, but also introduces substantial risks when model outputs compress large volumes of input into apparently coherent summaries (Small et al., 2023). Under these conditions, minority and dissenting positions are especially vulnerable to omission or underrepresentation. Zhu et al. (2025), benchmarking 18 LLMs across 3,000 participants, confirm this empirically: LLM-generated summaries systematically produce higher representativeness scores for majority opinions than minority ones, meaning the appearance of a representative output routinely conceals whose views were actually captured.

By contrast, designs that preserve contestation can reduce division without suppressing difference. Tessler et al. (2024) demonstrate this with the Habermas Machine, where text-embedding analysis confirmed that both majority and minority critiques substantively shaped revised group statements, leaving participants less divided without suppressing dissenting positions, which effectively enables the mutual understanding that mass deliberation at scale can uniquely generate. However, these results come from controlled deliberative settings rather than binding international public consultations, so their applicability to large-scale global governance remains an open empirical question.

The abovementioned considerations inform the following proposition:

P3: AI-assisted deliberation strengthens mutual understanding when it preserves minority and dissenting frames alongside majority positions, but weakens it when consensus-optimising design suppresses diversity of citizen perspectives in the drive toward agreement.

2.3.4 Determination

In conventional collaborative governance, determination refers to the stage at which participants collectively reach conclusions such as procedural decisions, or substantive policy outputs (Emerson et al., 2012). In international policymaking contexts such as the C20, however, determination often does not work this way. Citizens’ inputs are received, processed, and translated into institutional outputs by a third party, the secretariat, the working group, or the policymaking body, through a process that is typically non-transparent to participants. This structural feature of global governance makes determination distinctive: the question is not whether participants reached an agreement but whether the translation from citizen input to institutional output was faithful enough for participants to recognise their concerns in what determination produced. The shared motivation dimension most at stake is commitment: when citizens cannot trace their contributions to the outputs the process generates, the consultation appears symbolic rather than substantive, and their willingness to engage in future rounds is undermined.

This accountability gap is well-documented. Smith (2009) argues that feedback mechanisms linking citizen input to policy outputs are a structural requirement of democratic innovation. Curato et al. (2017) identify feedback as one of the most consistently absent features of participatory governance design. Reviewing 23 ICT-enabled participation platforms across 17 countries, Peixoto and Fox (2016) find that while digital platforms have expanded policymakers’ capacity to receive citizen voice, most have not increased their willingness to act on it, revealing a persistent gap between contributing input and seeing it reflected in outputs.

AI tools can strengthen commitment at the determination stage by making this translation more visible. Congruency analysis can be used to assess whether the framing priorities of policy-facing outputs correspond to those expressed in citizen submissions (Vrydagh, 2022). When made publicly accessible, such analyses can give participants a way to evaluate whether their concerns were carried forward into recommendations, declarations, or other agenda-setting documents. The C20 case illustrates this empirically: automated congruency analysis reported in Ustenko et al. (2026) finds substantial divergence between the distribution of citizen frames in the survey corpus and the distribution of frames in the final G20 Declaration.

On the basis of these findings, the following proposition is advanced:

P4: In international policymaking contexts, AI-assisted participation strengthens commitment when citizens can recognise their frames in agenda-setting outputs, but weakens it when policy documents are not connected to the inputs citizens contributed.

3 Empirical illustration

3.1 Data

The empirical illustration addresses the initiation and discovery stages of principled engagement using the materials of the public consultation conducted by Civil 20 in 2023. The C20 public consultation attempted to address one of the standard criticisms levelled against participatory processes: that participation often remains socially unequal, with some groups underrepresented in ways that can affect perceived legitimacy (Allegretti, 2021; Christensen et al., 2015). At the request and with the support of the C20 Chair and steering committee, an International Grassroots Survey was conducted involving 5,152 participants from 64 countries in 24 different languages. Because one of the survey’s objectives was to include perspectives from underrepresented groups, coordinators used purposeful sampling. They enlisted over 400 social workers worldwide, who reached out to colleagues and service recipients in their professional settings. The survey documents the baseline participation metrics that AI-assisted initiation tools would inherit and need to address (motivating P1) and tests whether aggregate topic prevalence suppresses semantic differences across demographic groups (testing P2). Because the consultation itself did not employ AI tools, the analysis is diagnostic rather than confirmatory: it establishes whether the patterns anticipated by P1 and P2 are present in a large-scale multilingual consultation, not whether AI-assisted design would remedy them.

3.2 Method

The initiation stage was assessed using both participation counts and participation quality indicators. The first indicator of participation quality is response length, measured as the word count of each participant’s translated answer to the open-ended question about the main concerns/problems people faced in day-to-day life. Response length serves as a primary metric because methodological studies frequently assume that longer narrative entries represent higher-quality data (Andrews, 2005; Denscombe, 2008; Holland and Christian, 2009). Using a conservative heuristic, responses of five words or fewer are classified as minimally elaborated, providing limited scope for downstream interpretation (five words is roughly the length of “bad economy and no jobs,” and most responses shorter than this are single-word or fragmented entries; the mean length of response in a survey is 8 words). The second indicator is hard non-response to two sensitive survey items on income adequacy and the need for multiple jobs, defined as a blank response field, which is consistent with economic disclosure discomfort (Tourangeau and Yan, 2007). Together, these indicators are used to assess whether groups that were formally included in the consultation were also able to contribute in sufficiently substantive and legible ways.

The discovery stage is operationalised by applying topic modeling to the survey’s open-ended responses. After standard preprocessing (lowercasing, tokenisation, and removal of punctuation, numbers, and stop words), open-ended responses were analysed using Latent Dirichlet Allocation (LDA), a topic modelling algorithm that identifies recurring patterns of word co-occurrence and assigns each response a probability distribution over a set of latent topics. The number of topics was determined by maximising coherence scores and inspecting intertopic distance maps to ensure that the resulting topics were semantically distinct and non-overlapping. This procedure identified ten distinct topics, each labelled by the authors on the basis of its highest-probability keywords (for details see Supplementary Table S1). Subgroup framing differences within these topics were assessed through keyness analysis–a measure of relative keyword frequency that identifies terms statistically overrepresented in a given subgroup relative to the full corpus, with log-ratio used to estimate effect size (Gabrielatos, 2018). Framing comparisons were conducted pairwise between subgroups (e.g., Global South vs. Global North, younger vs. older participants), with reported terms required to meet minimum effect-size and significance thresholds together with a document-frequency criterion excluding idiosyncratic usage. Full keyword lists and effect sizes for the region, age, and vulnerability comparisons are provided in Supplementary Tables S2–S4.

3.3 Results

3.3.1 Initiation

The C20 survey engaged 1,076 (20.9%) participants from the Global South, 3,693 (71.7%) women, and 2,371 (46.0%) economically vulnerable participants. English was the largest single language of the survey (31.8%), but substantial participation also came from a major European language block (33.8%) and an East Asian cluster (10.7%), alongside a wider set of additional languages (23.7%), including Hebrew, Portuguese, and Turkish. Participants with tertiary education made up 73.5% of the sample, the 41–60 age group was the largest (45.4%). The participation of men, younger people, and those with lower levels of education was less visible. These metrics suggest meaningful access, but not equal representation.

What these metrics do not show is whether those invited to participate were able to participate on equal terms (P1). Participation quality was uneven both in terms of the response length and non-response patterns. 36.2% of all open-ended responses are classified as minimally elaborated. Overall, 12.5% of participants skipped at least one of the two economic questions (Table 1).

Table 1. Inclusion and participation quality by group based on the C20 India 2023 international grassroots survey (n = 5,152).

ᵃParticipants indicating lack of money, inadequate pay, or need for multiple jobs. 𝑼hinese, Japanese and Korean (CJK), Turkish, Arabic, Hebrew, Cyrillic-script participants. CJK translation compression may understate word counts. ᶜSpans two main regions: Spain (n = 383, median 7 words, minimally elaborated responses 40.5%) and Latin America (n = 284, median 4 words, minimally elaborated responses 61.4%).

Participation quality also varies systematically across groups. Global South participants produced a median of 6 words against 9 in the Global North, and 46.6% of responses from the Global South fell below the five-word threshold, against 33.5% of those in the Global North (Mann–Whitney U r = −0.203, p_bonf<0.001). In terms of language differences, Spanish-speaking participants show a minimally elaborated response rate of 49.8%. However, when split by region, Latin American participants (n = 284) produce a median of 4 words and a minimally elaborated response rate of 61.4%, while for participants from Spain (n = 383) the corresponding indicators are 7 words and 40.5%. The within-group effect size (r = −0.318) is the largest regional gap in the study, and it was invisible in the aggregate. Education reinforces the pattern of inequality through a different mechanism. Over 32% of primary-educated participants avoid sensitive economic questions, whereas for holders of a university degree the corresponding indicator is only 10.8% (Cramér’s V = 0.112, p_bonf<0.001).

The intersectionality of region and education reveals a more nuanced picture of inequality. Within the Global South, response depth declines sharply at each educational level. More specifically, the education gradient in response depth is largely a feature of the Global South rather than of the sample as a whole. Within the Global South, median response length falls from 7 words among tertiary-educated participants to 4 among secondary-educated and 3 among primary-educated participants, while minimally elaborated response rates rise from 41.8 to 59.6 to 74.2%. By contrast, within the Global North, secondary- and primary-educated participants both produce a median of 8 words and statistically indistinguishable rates of minimally elaborated response (39.9 and 33.3%; r = 0.002, p = 0.98). The disadvantage is not confined to primary education: Global South participants with secondary education (n = 208) also show elevated rates at 59.6%, nearly double the Global North tertiary-educated baseline of 31.6%. Taken together, these results show that although the consultation reached structurally disadvantaged groups, these groups did not equally express their concerns, which might have had a detrimental effect on dimensions of shared motivation, such as mutual trust and internal legitimacy.

Although the C20 India 2023 did not use AI tools to facilitate the public consultation, these findings identify concrete design problems that AI-assisted consultation in similar settings would need to address. The most immediate application is AI-assisted submission scaffolding: multilingual conversational agents can guide participants through the submission process with probing questions, clarification prompts, and structured follow-up (Tavanapour et al., 2019). This would help the most disadvantaged participants in terms of socioeconomic background as well as participants from different cultural and linguistic contexts produce more legible contributions. Accessibility features such as simplified question reformulation, voice-to-text input, and plain-language outputs can lower the expressive burden of structured consultation formats (Chemnad and Othman, 2024). However, caution is warranted in deploying these instruments, since translation and prompting systems trained predominantly on Global North corpora may misread context-specific frames in Global South responses as less coherent or less policy-relevant, thereby reproducing the very inequalities they were introduced to mitigate (Joshi et al., 2020).

3.3.2 Discovery

The discovery stage of the C20 public consultations refers to the sensemaking step in which citizens’ open-ended responses are organized into a structured set of priorities. To assess Proposition 2, comparisons were made across two forms of discovery output: aggregate topic prevalence across the full corpus and subgroup-level frame analysis based on differential keyword usage.

Figure 2 shows the hierarchy of the main topics that were extracted based on the algorithmic analysis of 5,152 responses to the question on main issues people face in their day-to-day lives. Ten distinct topics were identified: Income and Job, Poverty, Work-Life Balance, Mental Health, Cost of Living, Health and Housing, Basic Needs, Social Relations & Safety, Sense of Crisis, Purpose in Life (labels were selected by the authors and are based on the exact words used by participants). For instance, the Income and Job topic contains keywords such as “economy,” “job,” “money” and “fear.” The Work-Life Balance topic is represented by words such as “work,” “time,” “money” and “balance.” The Cost-of-Living topic captures “lack of money,” “family,” “rent,” and “high price.”

Job & Income and Poverty are the most frequently mentioned topics, which reaffirms the well-established notion of the prevalence of economic considerations. Specific attention was given by participants to the issues of Health and Housing, which are often subjects of contention at the level of national policies. At the same time, the analysis of responses indicates an emergence of a post-materialistic orientation (Inglehart, 2018), as evidenced by the presence of Work-Life Balance and Mental Health among the top five topics. The topic of Sense of Crisis features prominently among top-of-mind considerations, and its significance would have been probably underestimated if we used closed-ended questions, which are less suitable for capturing emotional import and often reflect pre-conceived notions of policy-makers and researchers.

Aggregate topic prevalence, however, does not by itself show whether these shared topics are framed similarly across participant groups. Focus was on group-specific keyword usage because differences in lexical emphasis can reveal divergent framings even when overall topic prevalence appears similar.

Keyword analysis reveals systematic differences in how various groups of citizens frame social issues with the regional axis producing the most pronounced divergence (Figure 3, Supplementary Table S2). Global South participants disproportionately emphasise structural and immediate economic concerns such as unemployment insecurity economy hunger and hardship. By contrast Global North participants more frequently address social disconnection, environmental awareness and other postmaterialist concerns: isolation, loneliness, connection, climate, materialism and lifestyle. Notably the word affordable appears among 60 participants from the Global North and not once in the South sample suggesting that the two groups engage with economic hardship through fundamentally different conceptual frames.

Demographic and socioeconomic dimensions also reveal differences in framing. The issues of psychological strain and wellbeing are top of mind for the age group of 21–40, evidenced by distinctive use of such terms as burnout, screen, emotion, awareness, and balance. Participants aged 61 + are more concerned with physical risks, as reflected in terms such as death, shelter, and medicine (Figure 4, Supplementary Table S3). University graduates tend to rely more on psychological and systemic vocabulary: depression, anxiety, inequality, racism, healthcare, homelessness; whereas secondary-educated participants are distinctive only in their use of transportation. Gender dimension also reveals significant divergence. Female participants use a distinctive cluster of terms organised around personal safety, care responsibility, and household-level vulnerability such as woman, aggression, safe, home, illness, parent, elderly, while no terms distinctive for male participants emerged under the same filters. Women’s distinctive framing around care and domestic vulnerability is consistent with documented gender inequalities in access to public services such as healthcare more broadly (Fakkel et al., 2026).

Participants in the economically vulnerable group are more focused on immediate physical and logistical constraints suggested by terms such as road, transportation, inability, and protect, accompanied by a notable co-occurrence of terms related to psychological strain (depressed). By contrast, concerns of the non-vulnerable group are expressed through more reflexive and evaluative terms: expectation, wrong, demand, opinion, inner (Figure 5, Supplementary Table S4). This distinction suggests that the most vulnerable groups bear a double burden of material constraint and emotional strain that aggregate topic prevalence metrics are least equipped to capture. They have the highest stake in policy outcomes and face the greatest risk of having their concerns overlooked when discovery stops at the topic level.

The findings highlight insights from subgroup-level frame analysis that aggregate topic prevalence alone cannot capture. Most computational text analysis applied to public consultation data stops at the topic level, reporting what citizens are talking about without asking how different groups talk about it differently. The keyness analysis reported here shows that this is a consequential omission: groups that appear to share a concern may be expressing fundamentally different experiences of it. These differences may disappear in aggregate counts but matter for whether discovery genuinely supports mutual understanding.

Other AI-applications may strengthen this approach specifically in a multilingual context. Unlike conventional topic models, which may produce different topic structures for each language, cross-lingual topic models learn a shared topic space that enables direct comparison of responses across languages (Bianchi et al., 2021). Where predefined analytical categories are available, supervised multilingual language models offer a complementary approach. Rather than discovering themes, these models classify responses according to an existing coding framework. Applied to open-ended survey responses on democracy across nine languages and ten countries, they achieved performance approaching that of human annotators, demonstrating the potential for scalable and comparable analysis of multilingual qualitative data (Dahlberg et al., 2026).

Mapped onto the four propositions, the initiation-stage findings in Section 3.3.1 provide direct diagnostic support for P1, and the subgroup-level frame divergence recovered through keyness analysis in Section 3.3.2 supports P2’s claim that aggregate topic prevalence obscures meaningful differences. P3 and P4 concern the deliberation and determination stages, for which the C20 survey offers no observable analogue; support for them remains theoretical, resting on the same underlying logic that AI-assisted design succeeds when it preserves difference and fails when it smooths it over.

4 Discussion and conclusion

This paper has argued that understanding when and how AI strengthens citizen participation requires moving beyond tool-level evaluation to the stage-specific conditions of the governance process. By grounding the analysis in collaborative governance theory (Emerson et al., 2012), the present study shifts attention to the relational conditions that make participation meaningful: trust, mutual understanding, internal legitimacy, and commitment. Additionally, the existing work on AI-enabled participation was extended to the more weakly institutionalised context of international policymaking (Hilbrich, 2022), where consultations are often one-off, non-binding, multilingual, and conducted at scale. In doing so, the paper also contributes to the literature on multilingual democracy by showing that linguistic diversity shapes not only who participates but how issues are represented and whether participants can recognise their concerns in final outputs (Verhasselt, 2025; Cabrera, 2024).

The empirical analysis warrants elevating participant entry to the first stage of principled engagement, a consideration Emerson et al. (2012, p. 11) treat as preceding their four elements rather than as one of them. It is at this stage that imbalances in the composition of participants and in participation quality, across regions, languages, and levels of education, create a substantial threat to shared motivation, and these are challenges AI tools should be specifically designed to address. Multilingual conversational agents can lower participation barriers at initiation, but only if they prioritise contribution quality over submission counts (Tavanapour et al., 2019) and account for the fact that many existing models are pretrained on dominant-language data and may therefore misinterpret contributions from underrepresented language groups (Joshi et al., 2020).

The dynamics of the discovery stage pose a further threat to shared motivation, arising from the way AI is typically applied at this stage in international policy consultation. Topic aggregation can flatten out the diverse perspectives of different groups, producing a topic distribution that is accurate on average while concealing the concerns of particular groups. Definition, which Emerson et al. treat as one of the four elements of principled engagement, is not performed by participants under these conditions. At the discovery stage its terminological component is settled by translation, preprocessing, topic modelling, and label assignment, rather than negotiated among those whose concerns are being described. This calls for approaches that preserve heterogeneity across groups, including keyness analysis (Gabrielatos, 2018), structural topic modelling (Roberts et al., 2014), and BERTopic applications that support subgroup analysis (Grootendorst, 2022). Additionally, multilingual contexts require AI-enabled methods that support cross-lingual comparison, either by learning a shared topic space across languages (Bianchi et al., 2021) or by classifying responses into predefined categories using multilingual language models that approach human-level annotation performance (Dahlberg et al., 2026).

The deliberation and determination stages are treated here theoretically, and both propositions require empirical testing in settings where data on structured discussion and policy formulation are available. Even at this stage, the existing literature supports testable conditions. AI facilitation platforms can extend deliberation to mass participation, but only when designed to keep minority frames visible rather than optimising for apparent consensus (Zhu et al., 2025; Tessler et al., 2024). AI-assisted summarisation can close the accountability loop at determination, but only when it produces traceable links between citizen input and final policy documents (Peixoto and Fox, 2016).

The analysis of the C20 India 2023 consultation also produces empirical observations specific to this international policy process. Although the consultation reached many groups typically excluded from formal policymaking, participation quality varied systematically along regional and demographic lines in ways invisible to aggregate metrics. Moreover, shared topics did not imply shared meanings across communities: Global South participants disproportionately invoke structural and immediate economic concerns such as unemployment and insecurity, whereas Global North participants more frequently address issues such as loneliness, isolation, and climate. These are not failures of the C20 consultation design specifically, but structural features of large-scale digital consultation that AI-assisted design must explicitly address, by moving from purely efficiency-oriented to equity-oriented systems.

Several limitations qualify the claims made in this paper. The dimensions of shared motivation, trust, mutual understanding, internal legitimacy, and commitment, were not empirically measured. Their dynamics are inferred from patterns in participation quality and frame divergence rather than observed directly. Second, the C20 India 2023 consultation did not deploy AI tools, so the analysis cannot empirically test the efficacy of specific AI design solutions in mitigating the threats identified above. It establishes only the conditions and challenges that AI-assisted design in this setting would need to address. Longitudinal and comparative designs would allow stronger tests of the propositions formulated here, by measuring how shared motivation evolves across consultation rounds and evaluating whether specific AI-assisted interventions improve participation quality and frame preservation.

A further limitation concerns the scope of the framework itself. As noted in Section 2.2, several elements of Emerson et al.’s original model, including leadership and the other drivers of collaboration, and capacity for joint action, are treated as given rather than analysed here. More extensive theorising and empirical work is required to bring these elements back into the analysis, addressing questions such as the extent to which, and the forms in which, AI-assisted facilitation might compensate for weak leadership, and how consequential incentives can be strengthened rather than undermined by AI-assisted design.

More broadly, the paper suggests that the democratic value of AI in global citizen participation lies not in its capacity to process more voices more efficiently, but in its capacity to preserve the differences that collaborative governance must be able to hear, interpret, and carry forward.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation of this work the author(s) used ChatGPT-5.2 in order to streamline and grammar check written sections, and to debug analytical syntax. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpos.2026.1842794/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Shefali Virkar, Vienna University of Economics and Business, Austria

Reviewed by: Ernesto Vivares, Latin American Faculty of Social Sciences Headquarters Ecuador, Ecuador

Luciano Barbosa De Lima, University of Malaga, Spain