Beyond Polarization: Opinion Mixing and Social Influence in Deliberation

Mohak Goyal, Lodewijk Gelauff, Naman Gupta, Ashish Goel, Kamesh Munagala · 2026

Read Beyond Polarization: Opinion Mixing and Social Influence in Deliberation by Mohak Goyal, Lodewijk Gelauff, Naman Gupta, Ashish Goel, Kamesh Munagala (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-14221.txt.

Introduction

Mohak Goyal and co-authors argue deliberation research misses a key dimension of opinion change in 2026. Deliberative processes are often judged by polarization alone. Polarization can rise or fall as people talk. That lens misses whether deliberation reshuffles who agrees with whom.

Opinion mixing is their name for that reshuffling dimension. A process might move everyone in parallel, preserving the pre-deliberation rank ordering. Or it might scramble alliances while averages stay put. The team calls mixing a more diagnostic measure of social influence in deliberation. Rank shifts tell a tale that mean views hide. Two talks can end with the same mean but a quite distinct map of who stands with whom.

Catalog record

CREATORMohak Goyal, Lodewijk Gelauff, Naman Gupta, Ashish Goel, Kamesh Munagala
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-14221

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Beyond Polarization: Opinion Mixing and Social Influence in Deliberation Mohak Goyal, Lodewijk Gelauff, Naman Gupta, Ashish Goel, Kamesh Munagala Abstract Deliberative processes are often discussed as increasing or decreasing polarization. This approach misses a different, and arguably more diagnostic, dimension of opinion change: whether deliberation reshuffles who agrees with whom, or simply moves everyone in parallel while preserving the pre-deliberation rank ordering. We introduce opinion mixing , measured by Kendall’s rank correlation ( τ \tau ) between pre- and post-deliberation responses, as a complement to variance-based polarization metrics. Across two large online deliberative polls spanning 32 countries (MCF-2022: n = 6,342 n=6{,}342 ; MCF-2023: n = 1,529 n=1{,}529 ), deliberation increases opinion mixing relative to survey-only controls: treatment groups exhibit lower rank correlation on 97 % 97\% and 93 % 93\% of opinion questions, respectively. Polarization measures based on variance tell a more heterogeneous story: controls consistently converge, while treated groups sometimes converge and sometimes diverge depending on the issue. To probe mechanisms, we link transcripts and surveys in a third event (SOF: n = 617 n=617 , 116 groups) and use LLM-assisted coding of 6,232 discussion statements. Expressed support in discussion statements strongly predicts subsequent group-level opinion shifts; this correlation is amplified by justification quality in the statements but not by argument novelty. To our knowledge, we are the first to observe how different notions of argument quality have different associations with the outcome of deliberation. This suggests that opinion change after deliberation is related to selective uptake of well-reasoned arguments, producing complex patterns of opinion reorganization that standard polarization metrics may miss. 1 Introduction Understanding how individuals or groups of people change their viewpoints upon deliberation is a central question in computational social science. Theories of deliberative democracy argue that well-structured discussion — where participants exchange reasons, justify their positions, and encounter diverse perspectives — produces better-informed and more reflective opinions. Fishkin’s framework of Deliberative Polling institutionalized this vision by combining small-group deliberations and plenary expert panels with pre- and post-surveys to measure shifts in opinion ( 16 ) . A substantial literature has examined whether deliberation may alleviate polarization ( 18 ; 8 ) or exacerbate it ( 37 ; 40 ; 25 ) , depending on group composition and deliberative context, typically by tracking changes in group means or variance of opinions. Yet polarization and mean shifts, while important, capture only part of what deliberation might accomplish. Consider two convergence scenarios that are indistinguishable by just observing the opinion distribution: in the first, every participant shifts toward the group mean by the same proportion, preserving the rank order of who holds more or less extreme views; in the second, participants genuinely reconsider their positions, leading some who were initially skeptical to become supporters and vice versa. Both scenarios reflect fundamentally different processes. The first is consistent with social conformity or regression to the mean; the second suggests that deliberation facilitated genuine opinion revision—that exposure to arguments and evidence caused individuals to update their beliefs in ways that reshuffled their relative positions. The same logic applies when opinions diverge. Consider two scenarios that produce identical increases in variance: in the first, every participant becomes more extreme in the direction they already leaned—supporters become stronger supporters, skeptics become more skeptical, amplifying pre-existing divisions. In the second, participants reshuffle: some former supporters develop doubts after hearing counterarguments, while some former skeptics become convinced by new evidence, and the individuals who end up at the extremes are not the same ones who started there. Both scenarios register as equivalent polarization under variance-based measures, yet they carry very different implications. The first reflects echo-chamber dynamics. The second, despite producing greater disagreement, reflects engagement: participants updated their views based on the discussion, and the resulting distribution, though more dispersed, emerged from genuine reconsideration. This distinction matters for democratic theory. If deliberation merely compresses or stretches opinions while preserving relative ordering, it may reflect conformity or entrenchment rather than reasoned persuasion. However, if after deliberation, a rank reshuffling occurs, then it may suggest that participants are genuinely engaging with the substance of the discussion rather than simply drifting toward the median or reiterating prior positions. We propose to leverage this rank structure to characterize opinion change beyond polarization. We quantify opinion mixing using Kendall’s rank correlation ( \tau ) between pre- and post-deliberation surveys (Definition 3 ), which measures how much the relative ordering of individuals’ opinions is preserved. A \tau of 1 indicates perfect preservation of ranks, while lower values indicate greater reshuffling. Critically, this metric is invariant to translation and scaling: a population that uniformly shifts in a certain direction, or whose variance shrinks or grows, will still show \tau=1 if no one changes position relative to anyone else. Rank correlation thus isolates the component of opinion change that opinion distribution-based measures cannot detect. We complement opinion mixing with two additional metrics from prior literature that capture distinct phenomena. First, we examine variance change , which measures whether the population opinion converged (reduced disagreement) or became more dispersed (increased disagreement) over the course of deliberation. Second, we assess softening of stances , which we operationalize as mean reversion, or the negative correlation between individuals’ initial opinions and their subsequent opinion changes; intuitively, this captures the extent to which those who started with more extreme views (higher or lower) tended to move toward the group mean. Together, these three metrics decompose opinion change into three parts: whether people moved relative to each other, whether the group as a whole became more or less dispersed, and whether those with extreme opinions softened their positions. These metrics characterize what changes during deliberation; we also investigate why such changes occur. Deliberative Polling events produce rich transcripts of small-group discussions, yet systematic analysis of these texts has remained limited by the prohibitive cost of manual coding. For the Shaping Our Future ( SOF ) deliberation (Table 1 ), where we can link each room’s transcript to the corresponding pre- and post-survey responses, we apply an LLM annotator rubric adapted from prior work ( 20 ) to annotate each statement for (i) stance (expressed support or opposition toward the agenda), (ii) justification quality (examples and anecdotes to support the statement), and (iii) novelty (new information relative to the prior statements made in the deliberation group). We aggregate these annotations into group-level features and link them to pre–post opinion shifts at the deliberation group level, controlling for baseline opinions and topic fixed effects. This approach not only deepens our understanding of the mechanics of deliberation but also provides actionable insights for designing more effective deliberative polling events that foster better engagement and informed decision-making. Three Deliberation Events. Our datasets come from three deliberative polling events run on the Stanford Online Deliberation Platform: Meta Community Forum 2022 ( MCF-2022 ), Meta Community Forum 2023 ( MCF-2023 ), and Shaping Our Future ( SOF ). Table 1 summarizes their scale (participants, countries/languages, and number of small groups), format (sessions per participant), and survey design (number of opinion questions and the prevalence of “no opinion” responses pre/post). Both MCF datasets include a control group that completed the same pre–post surveys, enabling treatment–control comparisons. This group did not receive the briefing material

Read the complete text