DIEP Workshop
Andreas Flache: Full professor of sociology, Department of Sociology, Faculty for Social and Behavioral Sciences and the Interuniversity Center for Social Science Theory and Methodology, at University of Groningen.
Mirta Galesic: Cowan Chair in Human Social Dynamics at the Santa Fe Institute and Faculty member at the Complexity Science Hub Vienna.
Henrik Olsson: Group leader of the Collective Adaptation research group at the Complexity Science Hub Vienna
Mike Lees: Associate Professor, Computational Science Lab at University of Amsterdam
Han van der Maas: Professor of Psychological Methods, Department of Psychology at University of Amsterdam
Petter Tornberg: Assistant Professor in Computational Social Science at University of Amsterdam, senior researcher at University of Neuchatel, and Associate Professor in Complex Systems at Chalmers University of Technology
Lourens Waldorp: Associate Professor of Psychological Methods, Department of Psychology
Fernando P. Santos: Associate Professor at the Informatics Institute and group leader of Prosocial Dynamics Lab at University of Amsterdam
Joana Gonçalves de Sá: External faculty Complexity Science Hub Vienna
Petter Tornberg
Title: A Global Comparison of Political Discourse Quality Across Social Media Platforms
Social media platforms are often blamed for degrading public discourse, yet it remains unclear whether such effects are universal or shaped by platform design and political context. We present a global analysis of political discourse by examining over 2,700 political parties across 260 countries and five major platforms: Facebook, Instagram, TikTok, Twitter/X, and YouTube. Using computational measures grounded in deliberative theory, we find that reasoned, critical debate is uncommon across platforms. However, major differences emerge: some platforms systematically promote antagonism, while others discourage exposure to disagreement. Twitter/X—long treated as a ‘model organism’ in social media research—appears as a consistent outlier. These findings challenge prevailing assumptions about social media’s role in political life and reveal that platform design and national context jointly shape the quality of public discourse in fundamental ways.
Andreas Flache
Title: Modelling the emergence of affective polarization in social networks
As ethnic diversity increases in many societies, concerns about negative outgroup attitudes and interethnic affective polarization are growing. In social networks, affective polarization emerges from the complex interplay of a set of interconnected processes. I present a computational agent-based model of intergroup attitude dynamics, incorporating the effects of interpersonal intergroup contact, social influence by ingroup peers and changes in social network relations driven by attitudes towards the own and other groups. Using the ABM, conditions under which intergroup contact mitigates or exacerbates affective polarization are theoretically explored. Next, key assumptions of the theoretical ABM are connected to empirical data on the co-evolution of intergroup attitudes and friendship relations in diverse school classes. Starting from a baseline stochastic actor oriented model fitted to the data, the dynamics of network relations and attitudes in two different school cohorts are simulated. Simulations evaluate longer-term consequences of effects which external shocks could have on the relative strength of social influence on outgroup attitudes and of ethnic homophily in friendship selection. Findings indicate that peer influence, when intensified by an external shock, can over time magnify even small outgroup biases, leading to affective polarization. This dynamic is accelerated under strong ethnic homophily. A detailed analysis is provided of how these results emerge from the empirical context and the interacting processes included in the model.
Joana Gonçalves de Sá
Title: From Queries to Bias: A Bot-Based Audit of Search and Large Language Models in Election Contexts
Search engines (SEs) and large language models (LLMs) are increasingly shaping how individuals access information. While widely perceived as neutral, both technologies rely on opaque and dynamic algorithms, that are difficult to audit, raising concerns about their role(s) in reinforcing political polarization. I will describe a study that examines whether SEs and LLMs introduce or amplify political bias—potentially contributing to polarization—when responding to neutral election-related queries. We developed a large-scale, neutral and privacy protecting bot-based system (via web crawlers) that simulates users across different geographic and political contexts. We deployed these bots prior to the 2024 European Parliament and U.S. Presidential elections and directed them to conduct synchronized searches on SEs (Google, Bing, DuckDuckGo, Yahoo) and LLMs (ChatGPT, Copilot), collecting results for both election-related (e.g. ”who should I vote for”) and general queries (e.g. ”best restaurants near me”). We found that results varied significantly by location, especially for political queries, suggesting algorithmic tailoring that could create fragmented political realities. In the EU, SEs more frequently suggested right-leaning political content, with Google amplifying far-right sources via Wikipedia and official party links. LLMs also disproportionately mentioned right-wing entities when asked for voter guidance. In the U.S., while LLMs avoided direct endorsements, they emphasized Republican-prioritized issues more than Democratic ones, deviating from issue salience reported in national surveys, particularly when bots purported to be males. These findings suggest that algorithmically driven platforms may not only reflect but also reinforce existing political biases. I will discuss how these bots can be used to study algorithmic bias in general (we have also tested this system on the ongoing Israel-Palestine conflict) particularly in the context of increased gender polarization.
Mirta Galesic and Henrik Olsson
Title: Integrating Models of Belief Dynamics Abstract: Beliefs shape individual decisions and societal outcomes.
Understanding how beliefs form and evolve requires models that capture the interplay between personal cognition and social environments. Researchers have developed a variety of models often inspired by analogies to systems like ferromagnetism, epidemics, and evolution. These analogies provide conceptual mileage by highlighting parallels between belief systems and physical or biological processes, but they also carry conceptual baggage, especially when oversimplified assumptions are imported from the source domains. To help bridge and extend modeling of belief dynamics, we introduce the PES meta-model. PES structures belief dynamics around personal beliefs, expressed beliefs, and perceived beliefs of others. PES accommodates a wide range of social-psychological processes, including conformity, ego projection, authenticity, and social influence. The Networks of Belief theory, which connects individual belief networks with perceived and actual beliefs of others, is one specific implementation within PES. In addition, PES can incorporate a variety of classic models including the Voter, Ising, DeGroot, and bounded confidence models. PES offers a unified foundation for comparing mechanisms and grounding belief dymics in cognitive and social theory.
Fernando P. Santos
Title: The impact of link-recommendation algorithms on opinion polarisation
Online social networks are increasingly central in shaping human beliefs and behaviour. These are also prime spaces where humans co-exist with AI: algorithms to personalize contents and provide recommendations are pervasive in online platforms. Link recommendation algorithms are implemented to recommend new connections to online platforms users, based on supposed familiarity, similar interests, or the potential to serve as a source of useful information. These algorithms influence the evolution of social networks’ topology, yet their long-term impacts on human social dynamics remain unclear. In this talk, we will discuss how algorithmic link recommendations interplay with opinion dynamics, and the potential long-term impacts of such algorithms on polarisation. We will also discuss how such algorithms impact the visibility of opinions by different groups, and adaptations to restore fair recommendations.
Lourens Waldorp
Title: Designing effective policies for interventions to opinion formation
Fundamental to determining policy for opinion change in society is (1) to learn how opinions are formed, and (2) to learn what mechanisms remain the same as a result changes in society. In other words, it is not enough to know the mechanisms, we also need to know under which circumstances the mechanisms remain the same. For example, if the objective is to try to make people think positively of vaccination, then we have to first learn what causal mechanisms drive their opinion towards being vaccinated and then we need to know which types of information (e.g., news, anecdotes, scientific papers) and changes in society leave those mechanisms that drive their opinion the same. We use causal modelling to show the circumstances which allow for contextual invariances that predictions about which policies might be effective are likely. Here, we focus on several examples and determine exact conditions under which policies will be effective, and we show using simulations what kinds of distributional effects wee might expect when we either satisfy effective policy conditions or not.
Han van der Maas
Title: Modeling Psychological Systems with Ternary Spin Models: the case of zero
Psychological network models provide a promising alternative to traditional common cause theories, such as the g-theory of general intelligence and brain-based explanations of depression. These models, often formalized using the Ising model from statistical physics, have gained significant empirical support. However, the binary nature of nodes in Ising-type models presents a limitation, as many psychological datasets include responses with uncertain or neutral categories (e.g., ”don’t know” or ”not relevant”). Ternary spin models, such as the Blume-Capel model, overcome this constraint by incorporating a third node state that can represent such responses, enabling more nuanced scale representations. The resulting models exhibit more complex dynamics and provide new insights into research across a range of psychological constructs. We illustrate our approach with examples from three key subdisciplines of psychology. First, we introduce a ternary spin model for attitudes, extending the Ising attitude model. Next, we propose a unified framework encompassing both bipolar disorder and major depressive disorder. Finally, we present a novel ternary network model for understanding knowledge acquisition.
If you wish to register for this workshop, please send an email to m.t.pham@uva.nl.