Behavior at scale: a comparative analysis of diffusion, engagement, toxicity, and alignment patterns

Online platforms have transformed global information access, sharing, and consumption. They enable the large-scale study of discourse, diffusion, and the interplay of user preferences with algorithmic curation. However, by optimizing for attention, these systems influence the very behaviors researchers seek to understand. In parallel, as Artificial Intelligence (AI) systems become more deeply integrated into online platforms, behavioral patterns grow more complex. Within these dynamics, it is crucial to distinguish durable human patterns from contingent factors such as platform designs, which require broad longitudinal studies. This thesis combines large observational datasets with comparative modeling across regions, topics, demographics, and platforms to identify durable patterns versus design-driven variation. First, mapping 140M news articles across 183 countries, we show that a tightly interconnected group of countries disproportionately shapes the global agenda. Second, analyzing 12M comments about AI and vaccines across six platforms, we model engagement, revealing platform-specific interaction dynamics and distinct thematic emphases. Third, using 500M comments from eight platforms over three decades, we find that toxicity rises with conversation length, yet toxic language alone neither reliably expels participants nor escalates debates. Finally, across five datasets (220k annotations), we study demographic alignment of language models with human judgments and show that apparent demographic effects are inconsistent and confounded, where annotator sensitivity, agreement, and document difficulty explain more variance. Together this thesis provides empirically grounded guidelines for interpreting system effects and AI alignment claims.