This sub-project aims to closely examine the specific forms and underlying mechanisms of political, social, and cultural polarization generated by AI-driven platform technologies in online environments.

• RQ1: How do levels of ideological and affective polarization manifest in response to political and cultural events in online environments?
• RQ2: What structural linkages exist between the affective characteristics of content delivered through recommendation algorithms and users’ emotional responses? Specifically, how are the attributes of recommended content on the platform related to users’ affective reactions—and associated polarization indices—as measured using a large language model (LLM)-based emotion-response classifier?

To address the above research question, this study will first systematically document information on algorithmically recommended content on the platform using a sock-puppet auditing approach.

WP1 will then collect, at scale, user-generated materials that capture affective reactions—such as comments, posts, tweets, and YouTube comments—using tools such as Python, and analyze them with an emotion-response classifier grounded in a large language model (LLM). The LLM-based classifier will operationalize emotion along multiple dimensions, including hostility, contempt, empathy, and outrage, and will undergo a validation procedure to ensure reliability by cross-checking its outputs against human-coded annotations of the same materials.

Finally, based on the collected data, WP1 aim to quantify the temporal dynamics and diffusion structure of affective polarization using time-series methods and, ultimately, employ agent-based modeling (ABM) to model—within a causal inference framework—how recommendation algorithms amplify ideological and affective polarization in the broader public.