Dissertation, January 22, 2025

“Unless the spaces in a building are arranged in a sequence which corresponds to their degrees of privacy the visits made by strangers, friends, guests, clients, family, will always be a little awkward.” [10] ––– Unlike physical spaces where we have the capacity to move fluently between various environments, much of our online experience takes place in loud, crowded, and imposed public areas. This often hinders meaningful connections and understanding among individuals and deepens divides. Creating healthier, intentional public spaces, which involve mediating boundaries, social norms, and values, is crucial to addressing this challenge.

This thesis aims to establish that the intentional design of private and public spaces, along with bridges between them, can yield benefits in fostering healthier conversations. Our online experience is largely shaped by content that is filtered and ranked by algorithms according to centralized guidelines, which often do not align with a community's purpose. This thesis draws inspiration to help uphold communities to healthy standards. Ultimately, we intend to develop the knowledge necessary to advance content moderation strategies and governability for constructive communication across decentralized communities.

This dissertation seeks to collect and analyze historical social media data to identify the drivers of prosocial behaviors within online communities. Subsequently, it aims to design and test tools that allow communities to govern their interactions through content—ranking and moderation—aligned with their own speech norms. Additionally, it proposes new ways to enable communities to foster constructive communication within private spaces, as well as across communities into public spaces, assisted by language models.

In part one, we introduce and analyze a new dataset focused on community-centered content moderation from Reddit, where each community has set its moderation guidelines. We recovered over 230K removed posts, which are no longer available for public access. We examine over 19K communities defined by mission and over 60K speech norms and propose an empirically grounded norms schema for self-governed communities. We establish the relationship between prescriptive and restrictive behaviors as the guiding principle for the quality of conversation. Further, we analyze the correlation between intentionality in the design of a community and the quality of conversation measured by metrics like curiosity and compassion in interactions, revealing a high correlation between a community's mission statement and the quality of conversation.

Designing for communities to self-govern with their speech norms presents the challenge of shifting from a centralized, top-down approach to a distributed approach. While centralized methods offer efficiency in affordances when using highly accurate language models, they limit community members' participation in shaping how content is curated and ranked within their community. In part two, we explore the use of explainable language models for highly personalized content moderation and content uplifting.

Finally, in part three, we consolidate these learnings and propose a new social network space, Odessa. Odessa is a mobile app foundationally designed to foster communication and governance between distinct communities and promote human-human relationship building. It transitions from private spaces to public spaces with a shared mission, where communication is bridged with the assistance of large language models that help negotiate differing speech norms and values.