AI in Modern Society: From Computational Methods to Societal Consequences

Large Language Models (LLMs) are increasingly embedded in everyday online life. They serve as research tools and are integrated into systems through which people communicate, search for information, and coordinate with others. Yet, evidence remains limited on two related questions: whether LLMs provide methodological advantages for studying social phenomena, and how the integration into online environments changes practices relevant to civic participation.
This thesis addresses these questions in two complementary parts. First, we examine under which conditions LLMs can extend the Computational Social Science (CSS) toolkit for tasks such as synthetic data generation, text classification, and social simulation. Second, we explore how AI-assisted systems affect online discussion, information search, and coordination in social dilemmas.
Across three contributions, the first part shows that LLMs can extend existing CSS methods, but cannot be treated as general replacements for established methods or human annotation. Their performance and practical value depend on the task, the available data, and the computational resources. Through empirical comparisons, these studies rovide guidance on when LLM-based approaches offer meaningful advantages and when conventional methods remain preferable.
The second part demonstrates that integrating AI in online environments can reshape practices connected to civic participation. On social media, AI increases participation and content production, but at the same time make discussions perceived less informative, lower in quality, and less authentic. In information search, AI can improve task performance but also lead to reliance on generated answers rather than verification against original sources, leaving users vulnerable to misleading outputs. In social dilemmas, AI-generated messages can directly impact contribution behavior, although the effects weaken over repeated interactions.
Together, these findings show that the societal consequences of AI arise from model capabilities and how these capabilities are embedded in systems on the Web. With this thesis, we contribute methodological guidance for using LLMs in CSS and empirical evidence about their effects on online behavior. We also motivate further controlled, longitudinal, and ecologically valid research into increasingly AI-mediated environments.