UC Davis

Recommender systems shape how people encounter news, political information, and problematic content online. Their societal impact depends not only on what algorithms suggest but also on how users navigate and consume content, and on whether interventions can meaningfully reshape information diets. This dissertation brings together large-scale audits, a month-long field study with frequent YouTube users, and bottom-up interventions to advance three contributions.First, we conduct systematic algorithm audits with trained sock-puppets to measure YouTube's recommendation dynamics at scale, focusing on ideological congeniality, potential radicalization, and exposure to problematic channels. We find that recommendations are often ideologically aligned—especially for right-leaning accounts—while exposure to explicitly problematic channels is limited in volume yet non-trivial in reach. These patterns are consistent with platforms' optimization for engagement and personalization. Second, leveraging granular behavioral data from frequent YouTube users, we disentangle algorithmic from user-driven contributions to actual content consumption. Algorithmic pathways (homepage, sidebar, and search) account for the majority of congenial political consumption, with recommender ranking elevating like-minded content. By contrast, democratically problematic content is disproportionately accessed via user-driven pathways such as channel browsing and off-platform links. Third, we design and evaluate interventions that steer recommender ecosystems without platform cooperation. In a month-long randomized experiment, where we unobtrusively seeded user watch histories histories with verified, ideologically balanced news increased both recommendations to and consumption of news content and enhanced ideological diversity without reducing engagement. Together, these studies provide a unified framework for auditing, attributing, and steering platform ecosystems toward healthier and more diverse information environments.