Modelling opinion dynamics in complex social systems: from agent-based to AI-powered simulations

This thesis investigates opinion dynamics in complex social systems by combining methods from statistical physics, network science, and artificial intelligence. The central aim is to understand how individual behavior, social structure, and computational models can be used to model social systems and predict collective outcomes, and how these insights can be applied to real-world scenarios. Specifically, this thesis focuses on how to model opinion formation and describe opinion shifts in the population, on how to probe opinions through new surveys that exploit network effects, and on how to simulate human agents with new AI tools. We begin by showing how a simple modification of the classical voter model — introducing a latency period after each opinion change — fundamentally alters its behavior. Instead of drifting randomly toward consensus, the system exhibits persistent oscillations, offering a possible explanation for recurrent opinion swings observed in electoral and social contexts. This result highlights how small changes in individual rules can generate qualitatively new forms of collective dynamics. We then investigate how structural features of social networks affect the way opinions are measured. By analyzing the “social circle” polling method within a generative network framework, we show how network topology simultaneously biases these kinds of survey results and reveals the level of polarization in society. We develop an improved estimator that combines individual responses with structural information, and demonstrate its applicability on polling data from the 2016 U.S. presidential election. Finally, we extend the scope of opinion dynamics modeling by employing Large Language Models (LLMs) as artificial agents in online political conversations. Using Reddit discussions from the 2016 U.S. Presidential election as a testbed, we assess the ability of GPT-4 to reproduce human-like behavior. The experiments reveal both the realism and systematic biases of LLM-generated discourse, pointing to their potential as tools for simulating complex interactions but also raising concerns about manipulation and governance in digital environments. Moreover, we highlight a potential method for spotting LLM-generated texts inside social networks. Together, these studies show how the interplay of individual decision rules, network structure, and advanced AI systems can be captured within a unified modeling perspective. By bridging approaches from statistical physics, empirical data analysis, and artificial intelligence, this work contributes new theoretical mechanisms, practical methodologies, and critical reflections on the future of social simulation with physical approaches.