Investigating Social Polarization in Digital Networks by Combining Simulated Generative Agents and Human Users in Empirical Studies

• Applicant: Dr.-Ing. Tim Donkers
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 551054271

The rapid growth of online social networks has raised concerns about social polarization and ideological echo chambers. These are thought to be spaces where users repeatedly see their beliefs reflected without exposure to dissenting opinions, in part due to recommendation algorithms that determine what users see. While traditional research methods such as social network analysis provide a broad perspective, they do not allow us to analyze in detail how individuals interact with information and form their beliefs. Our project aims to fill this gap by combining experimental user studies with simulation technologies to understand the emergence and evolution of ideological positions in digital environments. Specifically, we will develop a simulation model that integrates both language processing and recommender systems to mimic real-world social networks as closely as possible. Existing open-source large language models will be refined using existing or self-created social network datasets to credibly mimic the communication behavior of human users. At the same time, current recommendation technologies will be used to simulate the spread of messages in the network. The resulting synthetic social network will then be made accessible via a user interface similar to that of conventional social networks. Based on this social network, we will conduct experimental studies with real people to gain empirical insights into the mechanisms that drive social polarization. Subjects will interact with the virtual agents through the interface described above, allowing us to observe and understand the interplay between social behavior, cognitive processes, and algorithmic influences. The pronounced control that we as researchers have over this network allows us to pose completely novel research questions that would be difficult to answer with the available repertoire of social network analysis, observation of user interactions in real social media, and isolated simulations alone. In particular, we want to use different recommendation variants that either narrow or widen the information corridor as experimental conditions of our empirical investigations in order to develop concrete strategies for promoting healthier online interactions.

• DFG Programme: Research Grants