At the heart of the framework lies a dual-threshold mechanism that governs how any two individuals influence each other. When the opinion distance between two agents falls below a lower threshold, assimilation dominates and the agents move closer together, a mathematical stand-in for the everyday experience of finding common ground. When the distance exceeds an upper threshold, repulsion takes over and the agents actively push their opinions apart, capturing the polarising effect of encountering views perceived as alien or threatening. Between the two thresholds lies a zone of non-interaction, in which people simply ignore one another. This tripartite structure echoes long-standing models in the opinion dynamics literature, but the authors integrate it into a continuous-opinion agent-based model in which every mechanism, every transition, and every final structure can be traced back to the same set of rules.

The genuinely novel step is what the researchers build on top of this interaction engine. Rather than treating phase transitions, steady-state structures, and multistability as separate research problems, as much of the existing literature has done, the team derives an empirical phase criterion directly from the initial distribution of neighbourhood opinion differences. This criterion is a single quantity whose sign indicates which of two competing forces will dominate the evolution of the system: contraction, in which opinions pull together toward consensus, or expansion and decoupling, in which opinions drift apart and the population fragments. In other words, by inspecting the statistical texture of opinion differences at time zero, the framework can forecast the qualitative fate of the entire discussion before a single interaction has unfolded.

The predictive power of this criterion was tested across a systematic grid of simulations spanning the model’s baseline parameter space. The results are striking for a social model: the empirical phase criterion achieved 86.7 percent agreement with the simulated end states, correctly identifying whether a given parameter combination would produce convergence or fragmentation in the large majority of conditions. For a field in which qualitative intuition often outruns quantitative prediction, this level of phase-sign agreement represents a meaningful benchmark, and it suggests that the initial microstructure of opinion differences carries genuine information about the macroscopic outcome.

Predicting whether a discussion will polarise is only half the problem. When polarisation does occur, the crucial question becomes how many distinct opinion clusters will survive at steady state. To answer this, the authors introduce a four-coefficient empirical prediction formula that estimates the steady-state cluster number under polarised conditions. The formula condenses the relevant model parameters into a compact expression whose output is the expected number of camps. Across the analysable conditions in the baseline grid, the prediction landed within one cluster of the observed modal value in 82.26 percent of cases. For moderators, platform designers, or anyone trying to anticipate how a controversy will shard, knowing whether to expect two camps or five is a substantial advance over simply knowing that fragmentation will happen.

The framework does not stop at a single deterministic forecast. Social systems are notoriously noisy, and the authors confront this head-on with a multistability correction term grounded in repeated-run statistics. When the same parameter configuration is simulated many times from different random seeds, the system does not always settle into the same final structure; different basins of attraction compete for the trajectory, and small stochastic fluctuations can tip the outcome toward one steady state or another. The correction term quantifies this variability, turning the framework’s predictions from point estimates into probabilistic statements that acknowledge the genuine contingency of social dynamics.

It is important, the authors stress, to read the term phase transition in an operational rather than a thermodynamic sense. In this framework, a phase transition refers to a reproducible change in the steady-state opinion structure that occurs across a narrow interval of a control parameter, analogous in spirit to the abrupt reorganisations studied in statistical physics but arising from social interaction rules rather than molecular forces. The simulations indeed reproduce the full catalogue of expected regimes: consensus, in which the population converges to a single shared opinion; bipolar polarisation, in which two opposing camps face each other; and multi-cluster coexistence, in which several distinct opinion groups persist indefinitely. The boundaries between these regimes can be sharp, and the empirical criterion locates them with the accuracy reported above.

Equally valuable is the study’s candour about the limits of prediction. Additional analyses show that stochasticity, initial conditions, finite-size effects, and competition among basins of attraction can all alter the realised steady-state structure, even when the phase criterion points in a clear direction. A population of one thousand agents may fragment differently from one of ten thousand; a slightly different initial histogram of opinions may steer the system into a neighbouring basin. These caveats are not weaknesses of the framework so much as honest descriptions of how complex social systems behave, and the authors are explicit that all of their evidence is generated within a single agent-based model. The framework should therefore be interpreted as a simulation-calibrated empirical account whose broader validity requires further testing across alternative empirical datasets and network settings before it can be applied confidently to real platforms.

Even with those caveats, the practical implications are considerable. Digital platforms increasingly face the challenge of anticipating how discussions will evolve, and the framework offers a concrete toolkit: measure the initial distribution of opinion differences in a thread or community, compute the phase criterion, and obtain an early warning of whether the conversation will converge or fracture; if fracture is likely, apply the four-coefficient formula to estimate how many camps will emerge. Because the entire chain of reasoning runs from measurable initial conditions through explicit interaction rules to predicted end states, the framework is also transparent in a way that black-box machine-learning forecasts of social behaviour are not. Every prediction can be traced to a mechanism.

The research also carries a conceptual message for the field. By demonstrating that interaction mechanisms, phase boundaries, cluster structures, and multistability can be handled within one coherent model, the study challenges the fragmented structure of opinion dynamics research and points toward a more integrated science of collective opinion formation. The work was supported by the 2025 Major Project of Sichuan Youth and Youth Work Research on mechanisms of youth social mobilization and group aggregation in cyberspace based on intelligent simulation, reflecting a policy interest in understanding how online communities coalesce and divide. As public discourse migrates ever more fully onto digital platforms, tools that can anticipate the tipping points between consensus and fragmentation, and that do so with quantified accuracy and stated limitations, are likely to become part of the standard analytical repertoire. This study is an early and rigorous example of what that repertoire may look like.

Subject of Research: Phase transitions and cluster formation in agent-based models of opinion dynamics

Article Title: A unified multiscale framework for phase transitions in opinion dynamics: integrating interaction mechanisms, establishing an empirical phase criterion, and predicting cluster structures

Article References: Jing, X., Hu, D., Xie, M., Liu, C., Zhou, X., & Lu, P. (2026). A unified multiscale framework for phase transitions in opinion dynamics: integrating interaction mechanisms, establishing an empirical phase criterion, and predicting cluster structures. Humanities and Social Sciences Communications. https://doi.org/10.1057/s41599-026-09420-0

Image Credits: AI Generated

DOI: 10.1057/s41599-026-09420-0

Keywords: opinion dynamics, phase transitions, agent-based modeling, polarisation, consensus formation, dual-threshold mechanism, opinion clusters, multistability, social simulation, computational social science, online discourse, empirical prediction

The research, published in Humanities and Social Sciences Communications, was conducted by Júlia Számely and Elisa Omodei of the Department of Network and Data Science at Central European University in Vienna, together with Júlia Koltai of Eötvös Loránd University and the Centre for Social Sciences in Budapest. The team linked traditional survey responses with donated Facebook trace data drawn from a representative sample of the Hungarian internet user population. This dual approach allowed them to ask two complementary questions at once: who, in socio-demographic terms, is more likely to engage with misinformation, and what patterns of online behaviour accompany that engagement?

Hungary offers an unusually compelling setting for such an investigation. Over the past decade and a half, the country has drawn international attention for democratic backsliding, including the concentration of political power, pressures on independent journalism, and a media landscape in which pro-government outlets dominate much of the information environment. In societies where the media ecosystem itself is contested, understanding which individuals end up engaging with misinformation becomes not just an academic puzzle but a question with direct consequences for democratic deliberation. Misinformation in such contexts is not a fringe phenomenon; it can be woven into mainstream channels of political communication, making individual-level vulnerability harder to isolate and study.

Methodologically, the study stands out for its use of donated social media data. Participants in a representative survey agreed to share traces of their Facebook activity, giving researchers access to behavioural records that would otherwise remain invisible. Self-reported survey measures capture what people say about their media diets, their trust, and their beliefs, but such reports are notoriously prone to bias. Trace data, by contrast, reveal actual behaviour: which pages users follow, which interest clusters they belong to, and how they move through the platform’s ecosystem of communities. By joining the two data sources at the individual level, the researchers could compare the predictive power of conventional demographic variables against that of lived online behaviour.

The central finding is striking in its simplicity. Among the socio-demographic factors the study examined, only age showed a consistent association with misinformation engagement. Other conventional suspects, such as education level, gender, or income, did not emerge as reliable predictors once the full picture was considered. This challenges a widespread intuition, common both in public debate and in parts of the research literature, that misinformation is primarily a problem of particular demographic groups. The Hungarian data suggest that knowing a person’s place in the social structure tells us far less about their misinformation engagement than knowing where they spend their attention online.

What did matter, strongly and consistently, was membership in specific interest clusters on Facebook. Online platforms like Facebook organize social life into overlapping communities built around shared interests, identities, and information sources. The study found that these clusters act as powerful filters of information exposure: some ecosystems amplify engagement with misinformation, while others are associated with lower levels of such engagement. In other words, the communities a user belongs to shape which content crosses their screen, and repeated exposure within a like-minded environment can normalize engagement with dubious content. A user embedded in an information ecosystem saturated with unverified claims encounters, and therefore has far more opportunities to interact with, misleading material.

This framing recasts misinformation engagement less as an individual failure of judgment and more as a structural property of online environments. The algorithmic curation and social reinforcement that define modern platforms mean that users do not confront the information landscape alone; they see it through the lens of their communities. If those communities consistently circulate and endorse false or misleading content, engagement follows almost as a matter of course. Conversely, users whose communities anchor themselves in more reliable information sources appear buffered against such exposure. The practical implication is sobering: interventions aimed purely at individual characteristics, such as media literacy training targeted at particular age or education groups, may miss the mechanism that matters most, which is the informational character of the communities people inhabit.

The study’s broader methodological argument deserves equal attention. The authors emphasize that structural socio-demographic factors should be studied together with online behaviour to provide a more complete understanding of what drives misinformation engagement. Neither data source alone tells the full story. A survey can situate a person within their social context, but it cannot see the communities they actually inhabit online. Trace data can map those communities with precision, but without survey linkage they remain detached from the social characteristics of the people behind the accounts. Only by combining the two can researchers disentangle the relative contributions of social position and digital behaviour. The Hungarian study demonstrates that this combination can overturn assumptions built on either source in isolation, since the apparent demographic patterns may dissolve once community membership is taken into account.

The research also carries lessons for how misinformation should be studied in countries undergoing democratic strain. Hungary’s changing media landscape, marked by shifting ownership structures and politically aligned outlets, makes individual-level engagement with misinformation especially salient, the authors note. In such environments, the line between misinformation and partisan mainstream content can blur, and engagement with misleading content may reflect the structure of available information sources as much as any personal disposition. Studies conducted in more pluralistic media environments may not capture these dynamics, which makes evidence from Hungary valuable for understanding how information ecosystems shape behaviour wherever media independence is under pressure. The findings resonate with a growing body of work suggesting that the supply side of misinformation, meaning the networks and pages that produce and circulate it, deserves as much scrutiny as the demand side of individual susceptibility.

The study was conducted online between February and June 2023, approved by the Research Ethics Committee of the ELTE Centre for Social Science in Budapest, and carried out with explicit informed consent from all participants, who received detailed information about the objectives of the research and data management before taking part, and who retained the right to request deletion of their data at any time. The work was funded through the CEU Open Access Fund and the Hungarian Academy of Sciences Lendület Program. Its publication in an open-access journal ensures that the evidence, and the methodological template it offers, are available to researchers and policymakers grappling with misinformation far beyond Hungary’s borders.

The takeaway for anyone concerned about the health of the online information sphere is that community is destiny, at least where misinformation is concerned. Age may mark who is more or less likely to engage with false content, but the ecosystems people join on social platforms appear to do far more work in determining what they see and how they respond. Efforts to slow the spread of misinformation may therefore achieve more by focusing on the structure of online communities, the sources they amplify, and the pathways through which dubious content reaches them, rather than on demographic profiles alone. As social media platforms continue to sort users into increasingly self-reinforcing interest clusters, the Hungarian evidence serves as a reminder that the fight against misinformation is fought not just in individual minds, but in the architecture of the digital spaces where those minds gather.

Subject of Research: Socio-demographic and online behavioural drivers of misinformation engagement in Hungary using linked survey and Facebook trace data

Article Title: Socio-demographic and online behavioural drivers of misinformation engagement in Hungary: evidence from linked survey and social media data

Article References: Számely, J., Koltai, J., & Omodei, E. (2026). Socio-demographic and online behavioural drivers of misinformation engagement in Hungary: evidence from linked survey and social media data. Humanities and Social Sciences Communications. https://doi.org/10.1057/s41599-026-09002-0

Image Credits: AI Generated

DOI: 10.1057/s41599-026-09002-0

Keywords: misinformation, Hungary, Facebook, social media, survey data, trace data, online communities, interest clusters, demographic factors, information ecosystems, democratic backsliding, computational social science

The difficulty lies in two competing traditions. Classical opinion dynamics models, such as the DeGroot model of 1974, the Friedkin-Johnsen model, and the Hegselmann-Krause bounded-confidence framework, describe how individuals update their views through explicit interaction rules. These models elegantly explain group-level phenomena like consensus formation and polarization, but they rest on simplifying assumptions: global interactions, homogeneous influence weights, or fixed neighborhoods. Real social networks, by contrast, are sparse and locally connected, and different neighbors exert very different degrees of influence on any given user. When the goal shifts from describing aggregate trends to predicting what a specific individual will believe next month, those idealized assumptions begin to crack.

Purely data-driven methods have the opposite problem. Continuous-time stochastic models such as SLANT and SLANT+ can process irregular event streams, while deep sequence models like LSTM and Transformers capture nonlinear patterns in individual histories, and graph neural networks incorporate network topology through neighborhood aggregation. Yet these approaches learn statistical correlations without any explicit formulation of social influence mechanisms. The authors point out that under observational noise, this lack of structural grounding causes substantial error accumulation over long prediction horizons, causing predicted trajectories to drift away from the actual dynamics of opinion change.

GA-ODE, short for graph attention with ordinary differential equations, belongs to an emerging middle ground sometimes called socially informed neural networks, which borrows ideas from physics-informed neural networks and Neural ODEs. The architecture has four cooperating components. First, each user receives a learnable embedding vector concatenated with their initially observed opinion, and a multilayer perceptron converts timestamps into continuous-time feature vectors. Second, a graph attention network computes unnormalized attention scores between a target user and each neighbor, normalizes them through a softmax function, and aggregates neighborhood information using a multi-head mechanism. This allows the model to learn differentiated aggregation weights, reflecting the fact that not all neighbors matter equally.

Third, inspired by the gating machinery of LSTM networks, the model employs a gated evolution unit to balance two opposing forces: individual opinion persistence and neighborhood assimilation. A forget gate, computed jointly from the temporal features and the user’s previous hidden state, determines how much of the historical cognitive state to retain. An input gate and a candidate cell state then control how much aggregated neighborhood information flows into the update, while an output gate regulates the contribution of the cell state to the final hidden representation. A self-attention module subsequently fuses the user’s hidden state with the temporal embedding, capturing the dynamic correlation between who a user is and when the prediction is being made.

The fourth and most distinctive component is a structure-aware ODE consistency constraint rooted in the Stochastic Bounded Confidence Model. Instead of a hard confidence threshold that forbids interaction between people holding divergent views, this probabilistic formulation assigns interaction probabilities inversely related to opinion distance, so individuals are more likely to engage like-minded peers but occasionally interact with those who disagree. The authors derive a continuous-time dynamical system from the discrete interaction rule, and because sampling discrete interaction partners is not differentiable, they apply the Gumbel-Softmax reparameterization trick to create a smooth, differentiable relaxation. The training loss then penalizes the squared difference between the time derivative of the predicted opinion trajectory, computed via automatic differentiation at randomly sampled collocation points, and the expected social influence term restricted to each user’s immediate neighborhood by an adjacency matrix mask.

The team evaluated GA-ODE against ten baselines, ranging from the classical Voter and DeGroot models to spatiotemporal graph networks like STGCN and T-GCN and socially informed rivals such as SINN, NPBC, and TT-SBCM. On three synthetic datasets engineered to produce consensus, polarization, and clustering dynamics, GA-ODE achieved accuracies of 0.892, 0.948, and 0.977 respectively. On a real-world dataset of political interactions gathered from the X platform between December 2023 and March 2024, containing roughly 10,883 messages from 526 users across 384 time steps, it reached an accuracy of 0.655. Compared with the strongest baseline, TT-SBCM, the improvements ranged from 1.5 to 3.7 percentage points, and two-tailed Welch’s t-tests across five random seeds confirmed the gains were statistically significant on every dataset.

Ablation experiments revealed why each piece matters. Replacing graph attention with uniform average-pooling aggregation dropped the macro-F1 score on the consensus dataset from 0.806 to 0.608, confirming that differentiated neighbor influence is essential. Removing the gating mechanism caused similar degradation, showing that opinion inertia cannot be captured by simply concatenating features. Perhaps most intriguingly, the topology-aware mask provided only limited benefits on clean graphs, where the graph attention encoder could already extract sufficient neighborhood information, but proved decisive under structural degradation. When 30 percent of edges were corrupted with spurious connections, GA-ODE maintained an accuracy of 0.621 while the mask-free variant fell to 0.551, and similar advantages appeared under missing edges and isolated nodes.

The robustness tests extended beyond network damage to the dynamics themselves. Real opinion change rarely follows smooth assimilation: sharp disagreement can trigger repulsive interactions driven by cognitive dissonance or the backfire effect, and sudden external events can jolt entire populations at once. When the team injected opinion-repulsion effects into the simulation, baseline models lost 4.6 to 7.3 percentage points of accuracy while GA-ODE lost only 2.3 to 2.8. Under external shocks delivered at specific time points, baselines dropped 3.2 to 5.2 points against GA-ODE’s 1.5 to 1.7. A t-SNE visualization further showed that the model’s latent representations form clearer, more discriminative clusters than those of the SINN baseline on the election data.

The authors are candid about the model’s limits. It assumes that social interactions remain static within the observation window, an idealization that real platforms routinely violate, and it is sensitive to sparsely represented opinion classes, which explains the gap between its accuracy and macro-F1 score on the imbalanced election dataset. Real-world validation also rested on a single platform and topic. Future work, they write, will pursue scalable graph learning and efficient ODE solving for large networks, dynamic graph modeling to capture the co-evolution of relationships and opinions, and broader evaluation across platforms and domains. Even so, GA-ODE offers a compelling demonstration that the century-old mathematics of social influence, when welded carefully onto graph attention networks and continuous-time dynamics, can make the messy evolution of human belief measurably more predictable.

Subject of Research: A hybrid machine learning model combining graph attention and neural ordinary differential equations to predict opinion dynamics on social networks

Article Title: An opinion evolution prediction model based on graph attention and ordinary differential equations

Article References: Liu, Q., Duan, L., Jia, M., & Li, Y. (2026). An opinion evolution prediction model based on graph attention and ordinary differential equations. Discover Artificial Intelligence, 6(1), Article 1474. https://doi.org/10.1007/s44163-026-02445-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02445-z

Keywords: opinion dynamics, graph attention networks, neural ordinary differential equations, social media, bounded confidence model, prediction, social networks, polarization, machine learning, computational social science, robustness, socially informed neural networks

The research, led by Jinghua Piao of Tsinghua University’s Department of Electronic Engineering with corresponding authors Fernando P. Santos, Yong Li, and James Evans, addresses a question that has grown urgent as LLM-powered autonomous agents proliferate across the internet. These systems can now generate social networks, hold conversations, and form shared or diverging opinions on political issues, yet the collective dynamics of such populations remain poorly understood. Individual chatbots have been probed extensively for bias and reliability, but far less is known about what happens when thousands of them interact with each other over extended periods, exchanging arguments and revising their views in response.

To find out, the team built simulations involving thousands of LLM agents built on five different backbone models: GPT-3.5, GPT-4o, ChatGLM, Llama-3, and DeepSeek-V3. Within these simulated societies, agents interacted through conversations guided by the underlying language models and updated their opinions over time as the exchanges unfolded. The scale matters. Classic computational social science experiments with human subjects are typically limited to dozens or a few hundred participants, but synthetic agent populations can be grown to sizes that make rare events and slow-moving structural changes visible. By running the same experimental logic across five distinct model families, the researchers could also ask whether the patterns they observed were idiosyncratic to one commercial system or a more general property of how current LLMs behave when placed in social settings.

The first major result concerns the shape of the networks the agents built for themselves. Rather than interacting uniformly at random, the agents spontaneously developed social networks with properties characteristic of human social networks, most notably homophilic clustering, the tendency of similar individuals to connect with one another. Homophily is one of the most robust findings in the empirical study of human relationships: people befriend people who share their views, backgrounds, and identities, and that sorting feeds back into what information they encounter. The fact that LLM agents reproduce this structural signature without being explicitly programmed to do so suggests that the social physics of similarity and connection may be latent in the statistical patterns these models learned from human-generated text.

The second major result concerns what happened to opinions over time. The collective opinions of the agent populations evolved in ways that exhibit behavioral patterns consistent with social phenomena and mechanisms widely discussed in empirical studies of human behavior and in classical opinion-dynamics models. Decades of research in this field, from averaging-based frameworks to bounded-confidence models in which individuals only influence those whose views lie within some threshold of their own, have catalogued the conditions under which societies converge to consensus, fragment into clusters, or split into diametrically opposed blocs. The LLM agent societies reproduced these canonical dynamics, and in particular gave rise to opinion polarization, the emergence of diverging, mutually opposed camps of views on political issues.

That consistency is the study’s most consequential claim, because it elevates LLM agent populations from a curiosity to a scientific instrument. If synthetic agent societies reliably mirror the dynamics documented in human data, the authors argue, they can serve as a valuable synthetic testbed for exploring hypothetical intervention strategies in networked LLM-agent systems. Testing anti-polarization interventions on real social networks is ethically fraught, slow, and expensive: platform-scale experiments touch millions of people, can influence elections and public health behavior, and are increasingly difficult for independent researchers to run. A synthetic society that behaves like the real thing offers a laboratory where interventions can be tried, failed, and refined at negligible cost and with no human subjects at risk.

Using this testbed, the team examined the effects of a range of network-level and individual-level interventions. At the network level, they studied the effects of promoting more diverse interactions, a strategy that echoes real-world proposals to break up filter bubbles by exposing users to viewpoints beyond their homophilous clusters. At the individual level, they explored reducing confirmation bias, the well-documented human tendency to favor information that supports existing beliefs and to discount information that contradicts them. Because the underlying agent architecture is fully observable and fully controllable, the researchers can trace exactly how a given intervention propagates through the network of conversations and reshapes the opinion distribution, something that is essentially impossible to measure at that resolution in human populations.

The findings arrive at a moment when the boundary between human and machine participation in online discourse is blurring. Autonomous agents already draft posts, reply in comment threads, and in some contexts interact with human users directly. If populations of such agents spontaneously form polarized, homophilous clusters when left to their own devices, then the deployment of large numbers of AI agents into shared information environments could, in principle, amplify the very dynamics that platforms and regulators are trying to dampen. The study does not claim that human polarization is caused by LLMs, but it does demonstrate that polarization is an emergent property of LLM-mediated social interaction, arising from the interplay of model behavior, conversation, and network structure rather than from any explicit instruction to divide.

There are also caveats worth keeping in view. The agents in these simulations are built on language models trained on vast corpora of human text, so their tendency toward homophily and polarization may partly reflect patterns absorbed from human writing rather than an independent social process. The researchers themselves frame the consistency with human data as a feature that makes LLM agents useful as a testbed, which implies that the agents inherit, rather than invent, the dynamics they display. Whether synthetic polarization tracks real-world polarization closely enough to predict the outcome of specific interventions on specific platforms remains an open empirical question, and the authors position their work as shedding light on opinion dynamics and demonstrating the potential of the approach, not as a finished policy tool.

Even so, the study marks a milestone for computational social science. It shows that the same discipline that spent decades building mathematical models of opinion flow can now grow living societies of language agents, watch them self-organize, and intervene in their dynamics with a precision that human experiments cannot match. As backbone models improve and agent populations grow, the gap between the synthetic testbed and the platforms it is meant to model will only narrow. The work, supported in part by China’s National Key Research and Development Program and the National Natural Science Foundation of China, suggests that the next generation of experiments on how societies split apart, and how they might be stitched back together, may begin not with millions of human users, but with thousands of machines talking to each other in the dark.

Subject of Research: Emergence of opinion polarization and homophilous network structure in simulated societies of large language model agents

Article Title: Emergence of polarization in networks of large language model agents

Article References: Piao, J., Lu, Z., Gao, C., Xu, F., Hu, Q., Santos, F. P., Li, Y., & Evans, J. (2026). Emergence of polarization in networks of large language model agents. Nature Communications. https://doi.org/10.1038/s41467-026-78228-y

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78228-y

Keywords: large language models, AI agents, opinion polarization, social networks, homophily, computational social science, agent-based simulation, confirmation bias, filter bubbles, GPT-4o, Llama-3, Nature Communications

The model rests on a well-established idea from social psychology known as optimal distinctiveness theory. People, the theory holds, are caught between two opposing needs: the desire to belong to a group and the desire to remain unique within it. Lean too far toward belonging and you lose your individuality; lean too far toward uniqueness and you lose your community. Most previous models of opinion dynamics have captured conformity, the pull toward agreement, but far fewer have built in this simultaneous push toward difference. The new work makes that push explicit, giving each simulated agent a reason to seek diversity among the people it interacts with most closely.

The second ingredient is what the authors call cognitive compression. Human beings are limited information processors, and one strategy for coping with complexity is to compress it: to reduce a sprawling landscape of nuanced positions into a handful of recognizable categories. In the model, agents evaluate how complicated the overall opinion environment is and prefer configurations that are simpler to represent. Technically, both the local diversity of an agent’s small social circle and the global simplicity of the entire opinion landscape are measured using Shannon entropy, the standard information-theoretic quantity that quantifies how much uncertainty or variety a distribution contains. High local entropy means a rich mix of opinions among one’s close contacts; low global entropy means the society as a whole is easy to summarize.

Each simulated interaction follows the same basic logic. Two agents meet, compare opinions, and decide whether to adopt each other’s views by weighing two objectives that pull in opposite directions. Adopting a different opinion can increase the diversity within the agent’s local group, satisfying the drive for distinctiveness, but it can also make the global opinion landscape more complex, violating the drive for compression. The agent’s decision therefore depends on a trade-off between maximizing local complexity and minimizing global complexity, both scored with the same entropy measure. It is a deceptively simple rule, yet when many such agents interact over many rounds, the collective behavior becomes strikingly rich.

The headline result is that polarization emerges naturally from this trade-off. Rather than converging to a single consensus or dissolving into random noise, the population organizes itself into distinct, heterogeneous opinion clusters, mirroring the fragmented but structured opinion groups observed in real survey data. This is significant because many existing models achieve polarization only by imposing assumptions that essentially build the outcome in, such as bounded confidence thresholds or pre-defined social networks. Here, the clusters arise from the bottom up, as an unintended consequence of individuals pursuing psychological goals that have nothing to do with division itself.

Perhaps the most surprising finding concerns what happens after the clusters form. In many classical opinion models, once the population settles into groups, opinions freeze: the dynamics effectively stop, and each cluster becomes a static bloc. In the compression-based model, individuals continue to adjust their opinions long after clusters have emerged. The result is ongoing variation both within and between opinion groups, so that the clusters remain alive and internally diverse rather than ossifying into monolithic camps. This continuous churn looks far more like real societies, where people drift within broad ideological families and the boundaries between groups shift over time, than like the frozen end-states of earlier simulations.

The computational experiments also reveal a sharp sensitivity to group size. Polarization, the authors find, appears when local group sizes are moderate, a regime the paper connects to Dunbar’s number, the famous estimate of roughly 150 stable social relationships that humans can maintain. When agents’ local groups are smaller than this moderate range, the population fragments into many small, weakly structured factions. When the groups are larger, distinct clusters struggle to form at all, and the opinion landscape blurs. The implication is provocative: the scale of our natural social circles, shaped by cognitive limits on relationships, may itself be a precondition for the kind of large-scale polarization that worries observers of modern politics.

The strength of cognitive compression acts as a second control knob on the collective outcome. When agents compress aggressively, simplifying the world strongly, the system becomes more unpredictable, with opinion trajectories that are harder to anticipate and group structures that shift in less regular ways. When compression is weak, the opposite occurs: the population settles into more consistent, stable group structures. In other words, how much individuals simplify the world around them changes not just whether polarization happens, but what kind of polarization it is, from rigid and predictable to volatile and erratic. This parameter could plausibly differ across cultures, media environments or historical moments, offering a lever for understanding why some societies polarize more violently than others.

Methodologically, the study is a showcase of what agent-based modeling does best. Polarization is a phenomenon that spans levels of description: it is produced by individual cognition, mediated by social interaction and visible only in population-scale patterns. No single discipline can capture all three at once, but an agent-based model can, by encoding psychological rules inside simulated individuals and letting sociological structure emerge from their interactions. The use of Shannon entropy as a common currency for both local distinctiveness and global simplicity is the technical keystone, allowing two very different psychological drives to be compared and traded off within a single decision rule. It is an elegant piece of mathematical bookkeeping that turns a vague psychological tension into a computable objective.

The broader lesson is that complex, realistic social behavior does not require complicated agents. Two simple rules, seek some diversity among your close peers and prefer a world you can easily summarize, are sufficient to produce opinion clusters, ongoing internal churn and a sensitive dependence on group size that echoes human cognitive limits. For researchers, the model provides a new tool for probing how changes in social scale or information environments might push societies toward fragmentation or consensus. For everyone else, it is a reminder that polarization may not be a bug introduced by technology, but an emergent property of minds trying, quite reasonably, to be both individuals and members of groups in a world that is too complicated to take in all at once.

Subject of Research: Agent-based modeling of opinion polarization driven by optimal distinctiveness and cognitive compression

Article Title: Opinion polarization from compression-based decision making where agents optimize local complexity and global simplicity

Article References: Dubovskaya, A., O’Sullivan, D. J. P., & Quayle, M. (2026). Opinion polarization from compression-based decision making where agents optimize local complexity and global simplicity. PLOS Complex Systems, 3(9), e0000115. https://doi.org/10.1371/journal.pcsy.0000115

Image Credits: AI Generated

DOI: 10.1371/journal.pcsy.0000115

Keywords: opinion polarization, agent-based modeling, cognitive compression, optimal distinctiveness theory, Shannon entropy, complex systems, Dunbar's number, social psychology, opinion dynamics, PLOS Complex Systems, emergence, computational social science

The motivation goes beyond simple data collection. In cybercriminal ecosystems, low-level identifiers such as channel handles are easily replaced, but higher-level semantic structures, including the language, products, and networks that define a campaign, tend to persist even as the surface infrastructure changes. This is why the authors argue for continuous observation rather than snapshot sampling. Archiving even short-lived communities supports the identification of illegal content under the EU Digital Services Act, enables the extraction of Indicators of Compromise for characterising the Crime-as-a-Service economy, and can ultimately assist the prosecution of cybercriminal activity. At the same time, many researchers currently depend on commercial Software-as-a-Service intermediaries such as TGStat or Telemetrio, which offer basic metrics but introduce opaque filtering, undermine corpus representativeness and reproducibility, and lock data inside proprietary silos without multi-dimensional enrichment.

Existing open-source tools each cover only part of this workflow. pytopicgram focuses on local extraction, preprocessing, sentiment analysis, and topic modelling; 4CAT provides persistent datasets, queued processing, and modular analysis processors; and TeleCatch offers a remote interface for filtered collection and message extraction. A more comprehensive framework called MST has been proposed in the literature, but because no versioned, publicly accessible source repository could be identified, the authors treat it strictly as related work rather than a direct competitor. In a capability comparison verified against public repositories and documentation as of August 2026, Atenea stands out as the only tool of the four that combines continuous monitoring, a persistent server-side message repository, named entity recognition, sentiment analysis, embedding generation, and, critically, true semantic retrieval over stored messages.

Architecturally, Atenea separates collection, processing, storage, retrieval, and exploration behind a common REST interface, with heavy inference decoupled from the core backend so that monitoring continues uninterrupted while services scale or evolve independently. Its data model revolves around three concepts: rooms, which are monitorable Telegram channels or groups; users, whether human accounts or bots; and seeds, external references such as invite links that define which rooms should be ingested. A dedicated Telegram Auth Manager registers and manages multiple API credentials, increasing effective request capacity and handling temporary restrictions. Two load-balancing strategies, one that splits work across credentials and one that groups resources by credential, keep collection resilient, and credential-to-room associations allow access recovery when a particular credential expires.

The processing layer relies on asynchronous, distributed ETL pipelines built from Telethon, Redis, Celery, and Celery Beat, triggered manually or by an automated scheduler. PostgreSQL serves as the transactional system of record, Elasticsearch powers lexical search and exploration through Kibana, and Qdrant stores vector embeddings for semantic retrieval. Downloaded media live outside the relational database in S3-compatible object storage, deployable locally via Garagehq, while PostgreSQL retains their provenance, metadata, hashes, and processing states. Database-level locking and configurable task concurrency prevent worker collisions across pipeline stages. Once collected, messages flow into a service-oriented enrichment layer offering named entity recognition, sentiment analysis, and embedding generation through independent microservices, including custom APIs, vLLM, or Ollama backends.

The enrichment pipeline itself is notably hybrid and language-aware. During ingestion, Atenea cleans message text and assigns an ISO 639-1 language code using the Efficient Language Detector library, discarding detections below a confidence threshold of 0.15. Before any model-based processing runs, regular-expression detectors extract structured indicators, including email addresses, URLs, Telegram mentions, hashtags, and cryptocurrency wallet addresses for Bitcoin, Ethereum, Dash, and Monero, with emoji tokenisation handled separately. These detectors preserve character offsets and operate independently of language. Model-based named entity recognition currently runs for English and Spanish using spaCy 3.8.14 with its large English and Spanish pipelines, mapping model-specific labels onto a canonical taxonomy. Adding a new language requires only installing a compatible spaCy model, registering its code, and redeploying, without changes to the enrichment API, task pipeline, or database schema.

To demonstrate the platform at scale, the authors describe a 2.5-year operational deployment that, as of a July 2026 snapshot, held more than 191 million messages from 6,412 monitored channels and groups, with history reaching back to 2016 and nearly 40 million extracted entities alongside 7.1 million inferred embeddings. A case study built on the DarkGram dataset, a research catalogue of cybercriminal Telegram channels, shows the full workflow in action. From 329 imported seeds, the populate pipeline resolved 207 monitorable rooms. An initial historical-recovery scan stored over 510,000 messages, peaking at 177,357 messages in a single hour, before settling into recurrent four-hour scans averaging roughly 128 new messages per day. The result was a tagged, reproducible collection of 574,432 messages spanning nearly nine years of channel history.

The enrichment results offer a descriptive profile of what automated extraction can reveal. Of the DarkGram messages, 378,905 qualified for English or Spanish model-based processing, and the hybrid extractor produced 175,021 unique entities across the full snapshot. Comparisons against the rest of the repository proved revealing: URL labels appeared at nearly identical rates in both collections, around 27 percent of messages, but PRODUCT labels were overrepresented in the cybercriminal collection at 4.7 percent versus 0.7 percent in the baseline, a pattern consistent with illicit marketplaces. References to people, by contrast, were far rarer in DarkGram, reflecting the baseline repository’s focus on disinformation and political-news channels. Media cataloguing recorded metadata for 147,646 attachments totalling 6.41 tebibytes of potential content, with opaque .bin files and compressed archives accounting for roughly 6.09 tebibytes, a distribution the authors note is consistent with packaged software and resources circulating in such channels.

Perhaps the most striking demonstration is the platform’s semantic capability. On external hardware equipped with two NVIDIA L40 S GPUs, a vLLM server running the Qwen3-Embedding-4B model generated 2,560-dimensional vector embeddings for more than 563,000 messages in roughly five hours, averaging 107,400 messages per hour with a peak of 35,224 in a single fifteen-minute bin. Stored in Qdrant and linked to their source messages, these embeddings allow researchers to retrieve content matching natural-language descriptions, such as credential theft or illicit marketplace activity, rather than relying solely on dates or keyword matches. Combined with statistical anomaly detection, which flagged two high-activity windows in the DarkGram collection, including one week in late 2025 with 11,005 messages and a z-score of 8.17, the system turns raw monitoring into a queryable analytical environment.

The implications reach well beyond cybersecurity. The authors position Atenea as a flexible backend for computational social science, journalism, and cyber threat intelligence, supporting workflows from network analysis to narrative detection, social listening during protests and crises, and the automated recovery of indicators such as IP addresses, domains, URLs, wallet addresses, and cryptographic hashes. Because the platform is released under the MIT license, with full documentation, a reproducible deployment capsule, and modular microservices that can be extended without touching the core, it lowers the barrier for any research group to build a long-running, auditable observation post on one of the world’s most consequential yet least-studied communication platforms. In an era when online communities can disappear in an instant, tools that preserve both the messages and their meaning may prove as valuable as the analyses they enable.

Subject of Research: An open-source platform for continuous Telegram monitoring, NLP enrichment, and semantic retrieval

Article Title: Atenea: An open-source platform for continuous telegram monitoring, NLP enrichment, and semantic retrieval

Article References: de Paz, A., & Arroyo, D. (2026). Atenea: An open-source platform for continuous telegram monitoring, NLP enrichment, and semantic retrieval. SoftwareX, 36, Article 103067. https://doi.org/10.1016/j.softx.2026.103067

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103067

Keywords: Telegram, open-source software, natural language processing, semantic search, cyber threat intelligence, named entity recognition, computational social science, disinformation, Crime-as-a-Service, vector embeddings, data collection, Digital Services Act

The research team, led by Cristian Candia of Universidad del Desarrollo and Northwestern University, together with Mohammad Atari of the University of Massachusetts Amherst and Nour Kteily and Brian Uzzi of Northwestern University, examined 1,621,147 observations spanning 13 socio-political topics. The corpus included 530,104 Twitter posts, 1,048,653 Reddit submissions and 42,390 posts from 8chan, giving the analysis an unusually broad cross-platform scope. All content was drawn from publicly accessible archives, and all results were reported in aggregate at the platform-by-topic level, with no user identifiers retained.

Central to the study was a methodological innovation in how moral language is quantified. Rather than simply counting moral words, the researchers used Distributed Dictionary Representations, a technique that scores word embeddings against an expert-validated moral dictionary. This approach captures the semantic proximity of any word to moral concepts, allowing the team to detect moral content even when it is expressed indirectly. Crucially, the method distinguishes two separate properties of a post: moral loading, the overall degree of moral relevance, and moral density, the concentration of moral content across the words used.

That distinction proved decisive. Using negative-binomial regression models appropriate for count-based engagement data, the team found that moral loading was positively associated with engagement across all platforms and topics, with effect sizes ranging from 1.12 to 9.07, all statistically significant. In other words, posts that were more morally relevant overall did attract more attention, confirming the well-established moral contagion effect. But conditional on that level of moral loading, moral density showed the opposite relationship: the more concentrated the moral vocabulary within a post, the lower the engagement, with coefficients ranging from −4.71 to −0.40.

The relationship between density and engagement was not simply linear. Instead, the models revealed a clear peak: engagement was maximized at a moral density of approximately 0.30. Below that threshold, engagement dropped by a factor of 2.28; above it, engagement fell even more steeply, by a factor of 2.78. The pattern held consistently across the three platforms, which differ markedly in culture, moderation norms and user demographics, suggesting the effect reflects a general property of how audiences process moralized messages rather than an artifact of any single online community.

The authors interpret the penalty through the lens of cognitive processing. Highly moral-dense posts may impose a heavier reading burden, triggering the kind of processing friction that previous work on fluency has linked to reduced persuasion and sharing. There is also the possibility of habituation: just as repeated exposure to extreme stimuli dulls emotional response, a wall of moral rhetoric may exhaust rather than mobilize readers. A mediation analysis using cognitive-processing language measures from LIWC was conducted to probe this mechanism, though the researchers are careful to frame it as exploratory rather than definitive proof of causation.

The findings arrive amid an ongoing scholarly debate about moral contagion. Earlier influential studies reported that each additional moral-emotional word in a tweet increased its diffusion rate, but a 2021 reanalysis in Nature Human Behaviour questioned the robustness of that effect, and a 2025 pre-registered replication and meta-analysis sought to pin down the true effect size. The new study offers a potential resolution: moral language may indeed promote engagement on average, but only within a bounded range. Studies that pool posts across the full density spectrum could obscure the penalty that kicks in at high concentrations, helping explain why different teams have reached different conclusions.

Methodologically, the study demonstrates the value of embedding-based measurement over simple dictionary counts. The team validated their Distributed Dictionary Representation approach against alternatives, including transformer-based sentence embeddings and the publicly released MoralBERT models, and the core results survived these comparisons. The replication package, built in R and Python with documented package versions, is publicly available on the Open Science Framework, though it deliberately excludes verbatim post text and user identifiers to prevent re-identification of individuals. The researchers note that the study involved retrospective analysis of public content only, with no intervention in human participants’ behaviour.

The practical implications extend to anyone who crafts online messages, from activists and journalists to political campaigns and platform designers. The data suggest a counterintuitive strategy: moral framing works best when it is present but restrained. A post that signals moral stakes with moderate density can ride the engagement advantage of moral relevance, while one that saturates every sentence with moral condemnation risks triggering the penalty. For platforms whose recommendation algorithms amplify morally charged content, the results hint that not all moralization is equal, and that saturation may be a signal of diminishing, or negative, returns.

Caveats remain. The analysis is observational, so it identifies associations between language patterns and engagement rather than experimentally established causes. Engagement metrics themselves are shaped by platform algorithms, and the topics studied were socio-political, leaving open questions about whether the same density peak applies to non-political moral discourse. Still, by separating how much moral content a post contains from how densely that content is packed, the study reframes a decade of moral contagion research around a single elegant idea: when it comes to moral language online, there can be too much of a good thing.

Subject of Research: The relationship between moral language saturation and user engagement on social media platforms

Article Title: Saturation of moral language predicts lower content engagement on social media

Article References: Candia, C., Atari, M., Kteily, N., & Uzzi, B. (2026). Saturation of moral language predicts lower content engagement on social media. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02560-y

Image Credits: AI Generated

DOI: 10.1038/s41562-026-02560-y

Keywords: moral language, social media, engagement, moral contagion, overmoralization, word embeddings, Twitter, Reddit, 8chan, negative-binomial regression, computational social science, Nature Human Behaviour

The research team set out to answer a deceptively simple question: who is actually represented in mobile location data, and how does that representation vary across space and demographic groups? To do so, the investigators compared the demographic and geographic composition of large-scale mobile location panels against authoritative benchmark populations drawn from the United States Census. By aligning the observed device counts with census figures across multiple spatial resolutions, from the national level down to counties and finer census geographies, the researchers were able to quantify where and for whom mobile location data over-represents or under-represents the underlying population. This benchmarking approach transforms a vague concern about bias into a measurable, mappable quantity.

The multiscale design is central to the study’s contribution. Many previous audits of mobility data focused on a single geographic level, typically the nation as a whole, where biases can average out and appear modest. But the new analysis demonstrates that national aggregates conceal dramatic local variation. A dataset that looks reasonably balanced at the country level may severely distort the picture in individual counties, particularly in rural areas, on tribal lands, and in communities with distinctive age or income profiles. Because many downstream applications, from retail site selection to emergency response planning, operate at precisely these local scales, the discrepancy between national fairness and local distortion carries real consequences.

Technically, the analysis relied on comparing device-level residence estimates with census counts of population by age, sex, race and ethnicity, income, and urbanicity. The researchers computed representation ratios, expressing the share of mobile devices attributed to a group or area relative to that group’s share of the benchmark population. Values above one indicate over-representation and values below one indicate under-representation. By computing these ratios repeatedly across scales and demographic dimensions, the team built a comprehensive portrait of sampling bias that could be decomposed into spatial components, sociodemographic components, and their interactions. This decomposition matters because spatial and demographic biases are intertwined: certain places are home to certain populations, so a geographic gap is often simultaneously a demographic one.

The findings confirm and sharpen long-standing suspicions about who is missing from mobile location data. Older adults, particularly those in the oldest age brackets, appear substantially less often than their share of the population would predict, reflecting lower smartphone adoption, different app ecosystems, and greater privacy sensitivity. Racial and ethnic minorities show uneven representation that varies by region, with some communities under-represented and others appearing at rates closer to parity. Income gradients emerge as well, with lower-income households less likely to contribute devices to commercial location panels, a pattern consistent with disparities in smartphone access, data plan affordability, and the types of applications that collect and monetize location information.

Geographically, the study finds that rural counties are systematically under-represented relative to urban cores, where dense populations, heavy app usage, and the economics of the mobile advertising ecosystem concentrate data collection. The magnitude of the rural gap is not uniform: it widens in regions with older populations and lower broadband penetration, and it narrows in peri-urban counties that blend rural character with metropolitan labor markets. The authors also document variation across states and metropolitan areas, suggesting that regional culture, state-level demographics, and the spatial distribution of technology industries all leave fingerprints on the composition of location panels. For analysts, the practical implication is that bias correction cannot rely on a single national adjustment factor.

Why do these biases arise in the first place? The study situates the problem in the mechanics of how commercial location data are produced. Devices enter these datasets only when users grant location permissions to applications that partner with data aggregators, and only when the resulting location pings pass quality filters designed to remove noise and protect privacy. Each step in this pipeline, from app adoption to permission granting to data brokerage, filters the population in ways that correlate with age, income, race, and place. The result is a dataset that is enormous, in the tens of millions of devices, yet still systematically skewed. Scale, the authors emphasize, is not the same as representativeness; a bigger biased sample is still a biased sample.

The implications ripple across the sciences that depend on mobility data. Epidemiological models that used location data to estimate contact patterns during the COVID-19 pandemic, for example, may have under-counted the movements of older and lower-income individuals, precisely the groups whose vulnerability and exposure dynamics differed most from the smartphone-owning mainstream. Transportation planners who calibrate travel demand models with mobile data risk designing infrastructure around the behavior of a demographically narrow public. Studies of segregation, access to services, and environmental exposure that infer activity spaces from location pings may mischaracterize the daily geographies of communities that are thinly sampled. The new bias maps offer these fields a diagnostic tool: before using location data for a given region or population, researchers can now check how well that region and population are actually represented.

The authors also point toward remedies. Weighting and reweighting schemes, in which devices are assigned statistical weights so that the weighted sample matches census benchmarks, can reduce aggregate bias, though the study cautions that such corrections struggle when entire groups are nearly absent from the data. Hybrid approaches that blend mobile data with traditional surveys, administrative records, or census microdata can anchor estimates to known population structure. Perhaps most importantly, the researchers advocate routine bias auditing as a standard step in any analysis using commercial location data, analogous to the data-quality checks that are customary in survey research. Transparency about representation, they argue, should become a baseline expectation for both data vendors and the scientists who consume their products.

As mobile location data continue to expand into nearly every corner of computational social science, this multiscale audit of the United States serves as both a warning and a roadmap. The warning is that the digital traces left by smartphones reflect the inequalities of digital access and technology adoption, and that these inequalities are written into the data at every geographic scale. The roadmap is a practical methodology for measuring those biases, mapping them across the country, and correcting for them where possible. In an era when decisions about public health, transportation, and resource allocation increasingly rest on algorithmic interpretations of human movement, knowing exactly who is missing from the data is the first step toward ensuring that data-driven decisions serve everyone, not just those who carry a trackable device.

Subject of Research: Spatial and sociodemographic sampling biases in mobile location data across the United States

Article Title: Characterizing the spatial and sociodemographic sampling biases found in mobile location data through a multiscale analysis of the United States

Article References: Embury, J., Dodge, S., & Nara, A. (2026). Characterizing the spatial and sociodemographic sampling biases found in mobile location data through a multiscale analysis of the United States. Scientific Reports. https://doi.org/10.1038/s41598-026-70982-9

Image Credits: AI Generated

DOI: 10.1038/s41598-026-70982-9

Keywords: mobile location data, sampling bias, human mobility, demographics, census benchmarking, spatial analysis, rural under-representation, digital divide, data representativeness, computational social science, Scientific Reports, United States

The study, conducted by Zander Weisman Mintz of the LBJ School of Public Affairs at the University of Texas at Austin and Ji Ma, now of Indiana University Bloomington and the University of Oxford, combines a large-scale systematic literature review with cutting-edge computational text analysis. Drawing on OpenAlex, an open scholarly database, the researchers built a corpus through citation snowballing that began with four foundational review articles and expanded forward through more than 14,900 candidate works. They then screened this mountain of material using Meta’s open-source Llama3 large language model, employing role prompting and chain-of-thought prompting to classify each candidate against six qualifying conditions. Validation against a manually reviewed sample of 200 works achieved 91 percent agreement. The final corpus comprised 3,268 works, published between the mid-1980s and May 2026.

To extract the latent thematic structure of this literature, the team applied BERTopic, a neural topic-modeling technique that leverages transformer-based document embeddings and clusters abstracts into coherent topics represented through class-based term frequency–inverse document frequency. The method outperformed conventional bag-of-words alternatives such as latent Dirichlet allocation and non-negative matrix factorization, particularly on the short texts typical of academic abstracts. Hyperparameters were optimized through random grid search to maximize topic coherence, and outlier reclassification plus manual relabeling refined interpretability. The result was a map of twenty distinct topics spanning the entire field, from healthcare AI and municipal waste optimization to algorithmic fairness and citizen participation in datafied democracies.

But topic modeling alone would simply produce another taxonomy. The study’s central contribution is a functional framework that organizes the literature around four practical functions of public governance: creating public value, delivering public services, responsiveness to the public, and protecting state–society relations. The authors deliberately rejected more familiar organizing axes. Sorting by application domain scatters shared governance problems like bias and accountability across every sector. Sorting by technology tracks capability but ignores the public stakes that make government deployment distinctive. Sorting by governance risk collapses the field into its pathologies. And sorting by public value alone absorbs distinctions a public manager most needs to see. The four governance functions, the authors argue, are the questions a public manager actually asks before deploying a system: does it create value, improve a service, change how the organization hears from citizens, and what does it do to the state’s relationship with the governed?

The framework’s dimensions are enriched by four theoretical lenses that recur throughout the scholarship: technological operationalization, which asks how AI fits specific public problems; function of government, which maps AI techniques onto sectoral activities; the policy-cycle framework, which locates AI within agenda-setting, formulation, implementation, and evaluation; and public governance, which foregrounds who is affected. Notably, the authors show that technological operationalization is not confined to the application dimensions but runs vertically through all four: a model’s opacity, its optimization target, and its hardcoded thresholds are design choices that propagate into accountability deficits and entrenched inequality downstream. Governance failures, they suggest, are often most cheaply addressed upstream at the point of technical design rather than through oversight bolted on after deployment.

The empirical findings reveal a field in dramatic flux. Output grew rapidly through the late 2010s, peaked at 627 works in 2023, and remained elevated through 2025. More striking is the compositional shift. The public value dimension, dominated by domain applications such as energy-efficiency forecasting and air-pollution monitoring, accounted for 50 to 77 percent of works through 2017 but declined to roughly 16 to 22 percent by 2024 and 2025. The state–society relations dimension, encompassing algorithmic fairness, ethics, and regulation, more than doubled its share, from around 12 to 23 percent in the mid-2010s to 52 to 54 percent in 2024 and 2025. The transition concentrates around the public release of ChatGPT in November 2022, which the authors use as a heuristic marker rather than a causal explanation, noting that generative AI diffusion, regulatory momentum such as the EU AI Act, and an ongoing accountability debate all converged in the same period.

The post-2022 pivot is stark in the numbers. Classical domain-application topics shrank from about 19 percent of the pre-ChatGPT corpus to under 5 percent afterward, while several topics essentially became new literatures: ethics and equity in educational AI grew from 2 works to 66, and AI ethics, accountability, and transparency in public service delivery surged from 9 to 291 works, a thirty-one-fold expansion. Semantic analysis of the twenty topics shows that the six state–society topics form a tight, mutually similar cluster, reading as a single coherent research conversation despite spanning multiple domains, while the public value topics spread widely across incompatible technical vocabularies. The framework thus partitions the field asymmetrically, and the authors argue this asymmetry is itself informative for anyone navigating it.

Equally important is where the literature is thin. The responsiveness dimension, concerning where AI can support the policy cycle and democratic feedback, carries only about 462 works across just two topics, making it the smallest dimension despite AI’s obvious relevance to agenda-setting, consultation, and evaluation. Intriguingly, public administration journals are already over-indexing here: responsiveness accounts for about a quarter of works in top-ranked public administration outlets, nearly double its corpus-wide share. The authors sketch four research directions for this space, including how large language models reshaping the information environment of agenda-setting, whether AI-mediated consultation strengthens or hollows the link between citizen voice and government decision, what happens when optimizing tools are embedded in satisficing policymaking processes, and whether AI-driven evaluation genuinely closes the feedback loop or merely produces dashboards.

The framework also doubles as a practical diagnostic for public managers weighing a specific AI system. Each dimension supplies a question and a concrete check: whether the system creates value the public can recognize, whether it improves a service without eroding recourse to a human, whether it widens or narrows the channel through which citizen preferences enter decisions, and whether affected citizens can learn of, understand, and contest its output while responsibility remains attributable. Applied retrospectively to the City of Chicago’s 2016 predictive policing heat list, a system later judged less effective than a simple most-wanted list, the diagnostic surfaces failures on three of four questions, illustrating how design-stage deficits that later evaluations documented could have been identified before deployment.

The study acknowledges its boundaries, including reliance on abstracts rather than full texts, the undercounting of grey literature and government reports, restriction to English-language scholarship, and the possibility of calibration drift in large-language-model screening. Yet within those limits, the contribution is a translation layer: a compact, function-based vocabulary that makes an interdisciplinary literature, previously scattered across engineering, information systems, and public administration, discoverable and comparable across the communities that study it in parallel. As governments worldwide grapple with deploying AI under democratic constraints, the researchers offer not a new theory of public administration but something arguably more useful right now: a shared map of what is known, what is contested, and where the next decade of scholarship and practice most urgently needs to go.

Subject of Research: A computational mapping of artificial intelligence research in the public sector using a functional governance framework and BERTopic modeling

Article Title: Towards artificial intelligence for the public sector: framing and bridging academia and practice

Article References: Mintz, Z. W., & Ma, J. (2026). Towards artificial intelligence for the public sector: framing and bridging academia and practice. Global Public Policy and Governance, 6(2), 124-156. https://doi.org/10.1007/s43508-026-00148-w

Image Credits: AI Generated

DOI: 10.1007/s43508-026-00148-w

Keywords: artificial intelligence, public sector, public governance, algorithmic fairness, BERTopic, systematic literature review, policy cycle, public value, ChatGPT, accountability, e-government, computational social science

The distinction matters because the two effects demand completely different responses. If opinions converge because agents influence one another through structured exchange, that is a dynamical phenomenon — something that depends on network topology, repeated contact, and the rules of communication. If opinions converge because every model shares the same training data and the same fine-tuning choices, that is a bias phenomenon, and no amount of network rewiring will fix it. Conflating the two leads researchers to draw false conclusions about social dynamics whenever they use language models to simulate human populations, a practice that has grown rapidly across computational social science.

Classical opinion dynamics models, from the DeGroot framework to bounded-confidence models such as Deffuant and Hegselmann–Krause, have long separated individual predispositions from interpersonal influence. A human agent starts with a prior position and then updates it as a weighted function of the neighbors’ positions. When all agents start from identical priors, any observed change must come from the interaction term. When agents never interact, any observed alignment must come from shared priors. The new study imports this clean separation into the world of language models, where it is far harder to achieve, because an LLM’s ‘prior’ is opaque and entangled with billions of parameters shaped by training corpora, alignment procedures, and decoding settings.

The methodological core of the work lies in designing controlled experimental conditions that isolate the two channels. In one condition, model agents are allowed to exchange opinions across multiple rounds, each agent seeing and responding to the outputs of others, so that interaction effects can accumulate. In a matched condition, agents are queried in complete isolation, with no exposure to each other’s outputs, so that any consistency in their answers reflects only the models’ intrinsic tendencies. By comparing the trajectories of these two conditions across topics, model families, and network structures, the researchers can estimate how much of the observed opinion dynamics is attributable to social influence and how much to shared bias.

The results carry an important warning for anyone using LLMs as synthetic participants. Because large commercial models are trained on broadly overlapping corpora and aligned toward similar helpfulness and safety objectives, they exhibit substantial baseline agreement: asked independently, they often cluster around the same positions on political, ethical, and policy questions. When such models are then placed in an interaction network, this shared prior acts like a strong external field, pulling the whole population toward the same attractor regardless of the communication structure. Apparent consensus in an LLM society can therefore be an artifact of homogeneity in the underlying models rather than an emergent product of deliberation — the opposite of what genuine social consensus formation looks like in human groups.

The study also quantifies when interaction does matter. Interaction effects become visible when agents begin from heterogeneous positions, when prompts are constructed to suppress the models’ default leanings, or when network structures channel information asymmetrically so that some agents act as hubs and others as peripheral listeners. Under these conditions, the trajectory of collective opinion diverges measurably from the isolated-query baseline, and classical dynamical concepts — anchoring, threshold effects, polarization into clusters — reappear in recognizable form. This suggests that the tools of decades of opinion dynamics research remain applicable to artificial agents, provided the bias floor is first accounted for.

For the broader scientific community, the findings arrive at a moment of intense debate about the validity of LLM-based social simulation. Several recent papers have shown that synthetic samples generated by language models can reproduce survey response patterns with striking fidelity, raising hopes for cheap, scalable, and ethically uncomplicated substitutes for human participants. Other work has cautioned that such fidelity is skin-deep: models reproduce the central tendencies of human populations while flattening minority viewpoints, amplifying majority biases, and failing to capture the contextual sensitivity of real respondents. The new disentangling framework gives this debate a sharper analytical instrument, allowing researchers to state precisely which portion of a simulated social outcome they are willing to trust.

The practical implications extend to platform governance as well. As AI agents increasingly populate recommendation feeds, comment sections, and automated moderation pipelines, the opinions they express are not neutral background noise; they actively shape the informational environment that human users experience. If those agents converge on uniform positions due to shared training biases rather than deliberative processes, they could function as an invisible consensus machine, nudging public discourse in directions no deliberative body ever chose. Understanding the bias-versus-interaction decomposition is therefore not only a matter of methodological hygiene for simulator designers but a governance question for the emerging mixed human–AI public sphere.

The study points toward concrete best practices. Researchers using LLM agents to model opinion dynamics should always run the isolation control: query each agent independently and establish the bias baseline before interpreting any collective behavior. They should diversify model families, prompt formulations, and initial conditions to prevent a single homogeneous prior from dominating the result. And they should report the decomposition explicitly, distinguishing influence-driven convergence from bias-driven agreement, so that downstream readers know whether an emergent consensus in silico says something about social process or merely about the models themselves. As artificial agents become permanent residents of our information ecosystems, tools like this one offer a way to keep the distinction between what agents say to each other and what they were built to say firmly in view.

Subject of Research: Disentangling interaction effects from training biases in the collective opinion dynamics of large language model agents

Article Title: Disentangling interaction and bias effects in opinion dynamics of large language models

Article References: Brockers, V. C., Ehrlich, D. A., & Priesemann, V. (2026). Disentangling interaction and bias effects in opinion dynamics of large language models. Nature Communications, 17(1), Article 10077. https://doi.org/10.1038/s41467-026-77340-3

Image Credits: AI Generated

DOI: 10.1038/s41467-026-77340-3

Keywords: large language models, opinion dynamics, AI agents, training bias, computational social science, agent-based simulation, social influence, polarization, consensus formation, Nature Communications, synthetic respondents, AI governance