Mass Online Deliberation

• Mass online deliberation is a form of large-scale, structured digital debate where public reasoning is mediated by platform design and algorithmic tools.
• It employs methodologies like threaded discussions, item-centered architectures, and social network analytics to maintain coherent and equitable debate.
• Its implications include enhanced civic engagement and collective intelligence, alongside challenges of information overload, participation inequality, and fragmented discussions.

Searching arXiv for recent and relevant papers on mass online deliberation and adjacent platform-design work. Mass online deliberation denotes large-scale, structured public discussions—often about policy or collective problems—hosted on digital platforms, where participants exchange arguments, replies, votes, and other interaction signals over time (Koursaris et al., 22 Apr 2026). Across the literature, it is treated not merely as high-volume participation, but as a socio-technical form of public reasoning in which architecture, interface, and algorithmic mediation shape who can speak, how disagreement is articulated, and whether collective discussion yields legitimacy, learning, or decision impact (Frappier, 2023). Research on the topic spans civic technology, deliberative democracy, computational social choice, argumentation systems, multi-agent simulation, and AI-assisted platform design. A persistent theme is that mass scale expands reach and issue emergence, but simultaneously introduces fragmentation, information overload, participation inequality, weak common ground, and tension between deliberative quality and representational breadth (Frappier, 2023, Khazaei et al., 2014, Revel et al., 19 Mar 2025).

1. Concept, scope, and distinguishing properties

Mass online deliberation concerns enabling very large numbers of people to reason together, give input, and make or influence decisions through digital systems (Davies et al., 2013). In one formulation, it involves large-scale, structured public discussions about policy or collective problems, requiring corpora with multi-turn arguments, replies, votes, and interaction patterns over time (Koursaris et al., 22 Apr 2026). In another, it is framed through “online assemblies”: socio-technical spaces in which people gather digitally to discuss, debate, and sometimes decide on political or public issues, with digital architecture acting as a regulatory force in the sense of “Code is Law” (Frappier, 2023).

Several distinctions recur in the literature. Relative to offline assemblies, online assemblies are distributed rather than local, asynchronous rather than synchronous, and text-based rather than speech-based; they also allow “virtually everybody” to join, while making it possible for many to remain “pure spectators” unless they write or vote (Frappier, 2023). Relative to generic online participation such as petitions or crowdsourcing, deliberative systems are expected to support debate and argument exchange rather than mere expression of preference (Frappier, 2023). Relative to small-group deliberation, mass online deliberation seeks to preserve reason-giving, reflection, and responsiveness while moving from mini-public scale to “crowd” scale through platform design, ML, and NLP (Shortall et al., 2021).

The literature also treats mass online deliberation as a site of collective intelligence. One survey defines collective intelligence as “the intelligence that emerges from local interactions among individual people” and focuses on direct interaction via verbal communication in environments such as discussion forums, Q&A fora, deliberation systems, and comment sections (Khazaei et al., 2014). This suggests that mass online deliberation is both a democratic procedure and a computationally analyzable interaction system whose emergent properties depend on communication patterns, argument content, sentiment, and downstream effects on participation and judgment (Khazaei et al., 2014).

2. Structural conditions: scale, asynchronicity, and platform architecture

The defining affordances of mass online deliberation are scale, temporal extension, and interface-mediated interaction. Online assemblies can involve “virtually everybody,” lower barriers across space and time, and connect citizens with each other and with government (Frappier, 2023). Yet as debates stretch over “days, weeks, or even years,” participants enter at different times, read different subsets of content, and rarely view the whole discussion, producing fragmentation and making common ground difficult (Frappier, 2023).

A recurrent design problem is therefore how to structure very large discussions so that they remain navigable and deliberative. Early work on Deme treated online deliberation as support for the core tasks of democratic meetings—collaborative drafting, focused discussion, and decision-making—through an item-centered architecture with group spaces, meeting areas, documents, polls, decisions, and comments linked bidirectionally to agenda items and text locations (Davies et al., 2013). This design shows a general pattern: flat lists of messages are inadequate for complex deliberation, whereas item-centered and document-centered structures can preserve coherence, especially when paired with explicit decision rules such as majority, approval, plurality, and consensus procedures (Davies et al., 2013).

Large-scale civic platforms display other recurrent architectural choices. The comparative analysis of online assemblies identifies three main interface zones: topic at the top, a voting or stance interface in the middle, and the deliberative space below (Frappier, 2023). The same work distinguishes forum-like patterns, in which arguments appear in threads, from poll-like two-column patterns that force users into “pros” and “cons” before posting, thereby emphasizing opposition and often moving platforms closer to survey tools than deliberative spaces (Frappier, 2023). Decidim Barcelona is notable for combining threaded comments with explicit alignment markers on first-level comments—neutral, positive, or negative—so that stance remains visible while disagreement can generate reply cascades (Aragón et al., 2017).

Hybrid architectures further extend the design space. The report on hybrid deliberation argues that dialogue-based participation increasingly combines online and offline components, citizen-led and government-led organization, and multiple institutional contexts (Zhang, 2023). Cases such as Decidim Barcelona integrate online proposals and debates with offline meetings, while the Global Assembly combines a 100-person core assembly with local assemblies worldwide (Zhang, 2023). This suggests that “mass” in practice often emerges not from a single giant forum but from a layered ecology of participation spaces, some open and large-scale, others small and intensive.

3. Core interaction problems: inequality, overload, representation, and quality

A central finding across the literature is that scale creates both democratic opportunity and deliberative pathology. On the opportunity side, online assemblies expand participation, permit issue emergence, and can channel citizen input toward city planning, participatory budgeting, or policy consultation (Frappier, 2023, Davies et al., 2020, Bojic et al., 2016). On the pathology side, mass online deliberation is repeatedly associated with information overload, fragmented attention, uneven voice, dominance by a small subset of participants, shallow aggregation logics, and weak links from discussion to institutional uptake (Frappier, 2023, Behrendt et al., 2024, Davies et al., 2020).

Participation inequality is a persistent concern. Deme sought to lower attendance barriers through asynchronous participation, email integration, and flexible spaces, but still recognized digital divides and usability complexity as democratic flaws (Davies et al., 2013). Reykjavik’s platforms revealed social skewing toward more politically active, university-educated, and higher-income citizens, even in a highly connected society (Bojic et al., 2016). COLLAGREE similarly showed that 75% of registered users never posted, while participation was demographically skewed and celebrity posts attracted roughly three times more replies than average posts (Zhang, 2023). This suggests that formal openness is insufficient for inclusiveness.

Information overload is treated as both cognitive and organizational. The survey of massive online dialogues notes that large threads and networks make it difficult to locate relevant arguments, motivating topic modeling, clustering, social network analysis, and quality assessment models as support tools (Khazaei et al., 2014). The adhocracy+ paper identifies redundancy, lack of structure, and cognitive difficulty in keeping track of many comments as signature problems of large-scale participation (Behrendt et al., 2024). The CMV-based study of persuasion adds that under cognitive overload, participants rely more on heuristic cues such as reputation, so influence can become skewed by ethos rather than by argument validity alone (Manzoor et al., 2020).

A related tension concerns deliberative quality versus representational breadth. Algorithmic ranking can improve civility or conversational quality but also suppress legitimate viewpoints. “Representative Ranking for Deliberation in the Public Sphere” formalizes this as a conflict between quality-oriented ranking and the representation of diverse voices, proposing justified representation as a constraint on top-$k$ comment selection (Revel et al., 19 Mar 2025). This suggests that mass online deliberation must be evaluated not only by tone or coherence, but also by whether significant cohesive groups are visible in the platform’s attention architecture.

4. Deliberative mechanisms and platform patterns

Despite heterogeneity across platforms, several recurring mechanisms define the state of the art.

Threaded discussion remains the dominant scaffold for exchange. It appears in forums, civic tech platforms, and proposal systems, enabling many users to join over time but also leading to long, fragmented reading paths (Frappier, 2023). When combined with alignment cues, as in Decidim Barcelona, threading can make disagreement legible and productive. The Decidim study operationalizes deliberation structurally through discussion cascades: width as a proxy for representation, depth as a proxy for argumentation, and the cascade $h$-index as a proxy for deliberativeness (Aragón et al., 2017). The key empirical result is that negatively aligned first-level comments are more likely to generate broader and deeper cascades than positive comments, suggesting that structured disagreement is a driver of deliberative decision making (Aragón et al., 2017).

Argument structuring and mapping are a second family of mechanisms. ODSG uses tree-structured online discussion in which posts are marked as ISSUE, IDEA, AGREEMENT, or DISAGREEMENT, complemented by graph-based summaries prepared by moderators (Zhang, 2023). ConsiderIt asks users to develop pro and con points one at a time and update a stance slider, making tradeoffs explicit (Zhang, 2023). Deme’s in-text commenting and item-comment architecture serve a similar purpose in document-centered collaboration (Davies et al., 2013). This suggests a general platform principle: mass deliberation benefits when contributions are indexed not merely by chronology but by argumentative role.

Mutual evaluation and clustering provide a third family of mechanisms. “System-Generated Requests for Rewriting Proposals” describes a system in which participants submit proposals and evaluate others’ proposals along two axes—understandability and agreement—using either a three-value scheme (“agree”, “disagree”, “don’t understand”) or more refined scales (Fenizio et al., 2016). Agreement patterns define a weighted graph over proposals with edge weights

$W_{AB} = \frac{|A \cap B|}{|A \cup B|},$

where $A$ and $B$ denote supporter sets; thresholding this graph yields clusters of compatible proposals (Fenizio et al., 2016). Quality or clarity then ranks proposals within clusters, while system-generated rewrite requests target either controversial but widely supported proposals or poorly presented but potentially valuable ones (Fenizio et al., 2016). A plausible implication is that large-scale deliberation can be organized not only through forums and voting but also through iterative proposal refinement driven by collective appraisal.

Voting and participatory budgeting constitute a fourth mechanism family. Better Neighborhoods lets citizens submit project ideas, after which officials assess cost and feasibility and citizens allocate a fixed budget across projects; in 2015 the budget was 300 million ISK, about 0.35% of the city’s total budget, with turnout declining from 8.1% in 2012 to 7.3% in 2015 (Bojic et al., 2016). Decide Madrid and Better Reykjavik both link online proposals to formal municipal response pathways, though the former is constrained by high support thresholds and the latter by demographic skew and thin deliberation (Davies et al., 2020). These cases show that mass online deliberation often interleaves deliberation and aggregation rather than replacing one with the other.

5. AI, automation, and computational mediation

Recent work pushes mass online deliberation beyond static interface design toward AI-supported and AI-generated systems. Three directions are especially prominent: automated quality assessment, AI-assisted facilitation, and synthetic simulation.

The first direction is automated quality assessment. AQuA defines a unified deliberative quality score from 20 indices spanning rationality, reciprocity, civility, and storytelling (Behrendt et al., 2024). For a comment $x$, the additive score is

$s(x) = \sum_{k=1}^{K} w_k \, f_{\theta_k}(x),$

where $f_{\theta_k}(x)$ is the adapter-model prediction for criterion $k$, and the normalized version is

$$s_{\text{AQuA}}(x) = 5 \cdot \frac{s(x) - s_{\min}}{s_{\max} - s_{\min}}.$$

Weights $h$0 are Pearson correlations between expert-coded criterion scores and non-expert perceptions of whether a comment is enriching or value-adding (Behrendt et al., 2024). This preserves multidimensional transparency while producing a single 0–5 scale suitable for ranking, monitoring, or recommendation (Behrendt et al., 2024).

A related study on the Stanford Online Deliberation Platform operationalizes contribution quality through four 1–5 criteria: whether a statement includes examples or anecdotes to support the speaker’s point, introduces novel ideas, builds on previous statements and the proposal, and raises points likely to improve following discussion (Gelauff et al., 2024). Using the average of other human annotators as ground truth, GPT-4 outperforms individual annotators, remains competitive with pairs, and is surpassed by triplets across all four metrics (Gelauff et al., 2024). The same paper shows that individual nudges after prolonged inactivity increase the likelihood of requesting to speak in the next 30 seconds by 65%, while contributions following nudges have quality similar to non-nudged contributions (Gelauff et al., 2024). This suggests that AI-mediated participation support can increase contribution volume without degrading local discussion quality.

The second direction is AI-assisted deliberation support. The adhocracy+ extension introduces a Comment Recommendation Module and a Deliberative Quality Module (Behrendt et al., 2024). The recommendation module uses a fine-tuned BERT stance detector, trained on X-Stance and synthetic data generated with Mistral-7B, to recommend a contrary-stance comment and invite a reply (Behrendt et al., 2024). The quality module computes AQuA scores and highlights the three highest-scoring comments as “top comments,” sorted to the top and marked visually (Behrendt et al., 2024). This operationalizes reciprocity and rationality directly within the interaction flow rather than only in post hoc analysis (Behrendt et al., 2024).

The third direction is synthetic simulation. CHORUS is an agentic framework for generating realistic deliberation data via LLM-powered actors with structured personas, memory, tools, and a Poisson process–based temporal model (Koursaris et al., 22 Apr 2026). Formally, actors are $h$1, each governed by a persona $h$2; shared history is

$h$3

and posting/action times are scheduled through independent Poisson processes with rates $h$4 and $h$5 (Koursaris et al., 22 Apr 2026). In a deployment on Deliberate with $h$6 actors over $h$7 minutes, the system produced 3–13 posts/minute and 6–25 actions/minute, with advocates most active and experts less frequent but longer-form, evidence-based contributors (Koursaris et al., 22 Apr 2026). Thirty expert participants rated the resulting discussions 4.6 for content realism, 4.1 for discussion coherence, and 4.3 for analytical utility on a 5-point Likert scale (Koursaris et al., 22 Apr 2026). This suggests that synthetic data can serve as an ethically easier, tunable complement to scarce real deliberation corpora.

A separate AI-mediated scaling strategy appears in Conversational Swarm Intelligence. CSI decomposes a large group into many small chatrooms of 4–7 people, linked by LLM-based Observer and Surrogate Agents that summarize local discussions and propagate salient ideas across rooms (Rosenberg et al., 2023). In a 48-person comparison between standard chat and the Thinkscape CSI prototype, CSI produced 51% more content, 37% less difference in contribution quantity between the most active and least active members, and significantly greater preference and perceived impact among participants (Rosenberg et al., 2023). A plausible implication is that “mass” deliberation may sometimes be better achieved through networks of small-group conversations than through a single large interaction space.

6. Evaluation frameworks and empirical findings

Evaluation in mass online deliberation spans content realism, coherence, participation rates, user satisfaction, policy impact, persuasion, representational guarantees, and contribution quality.

One cluster of evaluations focuses on realism and utility of data or systems. CHORUS, as noted, was evaluated by expert participants along content realism, discussion coherence, and analytical utility, with means of 4.6, 4.1, and 4.3 respectively (Koursaris et al., 22 Apr 2026). The Stanford platform assesses quality at the utterance level and shows that LLM ratings can support rapid experimental evaluation of interface interventions such as nudges (Gelauff et al., 2024).

A second cluster examines behavioral effects in live platforms. CSI’s experimental comparison uses contribution quantity and equality metrics, showing more balanced participation than standard chat (Rosenberg et al., 2023). Decidim Barcelona uses structural properties of reply trees, showing that negative comments are disproportionate initiators of deliberative cascades (Aragón et al., 2017). ChangeMyView enables unusually direct persuasion measurement through the delta system: among 1,026,201 debates over 7 years, the platform average persuasion rate is 3.5%, and having 10 additional reputation points causes a 31% increase in the probability of successful persuasion over the platform average (Manzoor et al., 2020). This demonstrates that ethos has causal force in online deliberation, above and beyond argument content, especially under cognitive overload (Manzoor et al., 2020).

A third cluster evaluates representation and ranking. The representative ranking paper defines justified representation for a highlighted set $h$8 of $h$9 comments by requiring that every cohesive group of at least $W_{AB} = \frac{|A \cap B|}{|A \cup B|},$0 users be represented (Revel et al., 19 Mar 2025). The platform objective becomes

$W_{AB} = \frac{|A \cap B|}{|A \cup B|},$1

On two Remesh sessions about campus protests, with $W_{AB} = \frac{|A \cap B|}{|A \cup B|},$2, GreedyCC-enforced JR reduced the proportion of unrepresented users from 18% to 5% for engagement ranking, from 15% to 4% for diverse approval ranking, and from 4% to 2% for Perspective API “bridging” ranking, while incurring only about 5–18% loss in objective score (Revel et al., 19 Mar 2025). This suggests that algorithmic representation guarantees are compatible with quality-oriented ranking in clustered preference settings.

A fourth cluster concerns institutional uptake and legitimacy. In Reykjavik, Better Reykjavik and Better Neighborhoods illustrate the importance of predefining citizen impact and maintaining proportionate budgetary and administrative commitment (Bojic et al., 2016). Better Neighborhoods scores highest on the participation ladder because it transfers some decision power over a small budget share to citizens, whereas Better Reykjavik’s main weakness is delayed feedback and lack of transparency when adapting ideas to legal and technical requirements (Bojic et al., 2016). In Decidim Barcelona, the local government publicly committed to accept the most-voted proposals, and 71–75% of proposals presented for the strategic city plan were accepted into the Municipal Action Plan, though many proposals generated no debate and deliberative quality declined over time (Zhang, 2023).

7. Normative controversies, limitations, and future directions

Mass online deliberation remains normatively and methodologically contested. One controversy concerns whether digital scale undermines the very qualities deliberation is meant to secure. Online assemblies often prioritize voting and binary positions, leading platforms to “appear more like survey tools than deliberative spaces” (Frappier, 2023). Two-column pros/cons designs may facilitate the exacerbation of opposing views and symbolically resolve debate through aggregation rather than common reason-giving (Frappier, 2023). The review “Reason Against the Machine” argues that the literature is largely focused on technical fixes for scaling up deliberation while neglecting more nuanced requirements of high-quality deliberation, especially across gender, culture, inequality, cognitive ability, and linguistic diversity (Shortall et al., 2021).

A second controversy concerns synthetic, algorithmic, and AI-mediated forms of deliberation. CHORUS provides ethically easier-to-share synthetic deliberation data, but the authors stress that synthetic does not equal real; such data may miss emotional nuance, trolling, harassment, and norm violations, and there are not yet ablation studies or comparisons against real deliberation logs (Koursaris et al., 22 Apr 2026). AQuA and GPT-based quality scoring offer scalable evaluation, yet they embed specific normative assumptions about what counts as good deliberation and may reflect biases of expert coding schemes or crowd perceptions (Behrendt et al., 2024, Gelauff et al., 2024). The adhocracy+ modules make AI interventions visible through popups and badges, but the paper does not yet resolve questions of explanation, bias, or participant control (Behrendt et al., 2024).

A third controversy concerns power, manipulation, and public-sphere distortion. The Indo-Pacific blogosphere survey shows that mass online deliberation is often entangled with authoritarian legacies, cybertroopers, commentator bloggers, and congressional bloggers, with state and party actors competing to steer discourse through propaganda, harassment, and digital campaigning (Akinnubi et al., 2023). This suggests that mass online deliberation is not merely a platform-design problem but a governance problem involving regime type, media control, and the boundary between authentic participation and orchestrated influence.

Several future directions recur across the corpus. Hybrid deliberation is proposed as the future direction for dialogue-based participation because it combines online scalability with offline trust-building, facilitation, and inclusion (Zhang, 2023). AI support is moving from summarization and moderation toward norm-aware micro-interventions—stance-aware recommendations, top-comment highlighting, quality dashboards, and synthetic simulations (Behrendt et al., 2024, Gelauff et al., 2024, Koursaris et al., 22 Apr 2026). Social choice methods such as justified representation bring formal guarantees of viewpoint visibility into ranking and curation (Revel et al., 19 Mar 2025). Swarm-like architectures offer a path to real-time, large-scale conversational deliberation without collapsing into a single overloaded thread (Rosenberg et al., 2023). At the same time, foundational challenges remain unresolved: how to secure representativeness, how to handle abuse without suppressing dissent, how to integrate outputs into binding institutions, and how to evaluate normative quality beyond popularity, civility, or local coherence (Shortall et al., 2021, Davies et al., 2020, Bojic et al., 2016).

Taken together, the literature portrays mass online deliberation as a design and governance project rather than a settled institutional form. Its success depends on whether digital systems can combine scale with reason-giving, inclusion with navigability, disagreement with civility, and algorithmic assistance with democratic legitimacy. The field has moved from generic comment sections and petition sites toward structured argumentation systems, participatory budgeting platforms, AI-augmented forums, representative ranking, and even synthetic deliberation simulators. Yet the central criterion remains unchanged: whether these systems enable heterogeneous publics to exchange reasons, contest proposals, and shape collective decisions under conditions that are not merely participatory, but genuinely deliberative (Frappier, 2023, Davies et al., 2013, Shortall et al., 2021).