Personalizing Content Moderation: How Middleware Shapes User Perception and Experience on Social Media

Abstract

Social media platforms increasingly determine what users see through opaque recommendation algorithms, reducing individual control and potentially exposing users to harmful or distressing content. Current moderation practices are centralized, reflecting the values of platform owners rather than diverse user needs.This thesis explores how middleware, positioned as a cross-platform system for individual content moderation on social media, shapes users' perception and experience. To address this, the thesis follows the Research through Design (RtD) approach, adapted for Human-Computer Interaction (HCI) by Zimmerman. This approach produces and reflects on design artifacts to generate research knowledge, rather than treating design purely as an output. Four artifacts were created, each building on insights from the preceding one. A discovery workshop (n=5) investigated the problem from multiple perspectives to inform the design in a user-centered way. Wireframes then conceptualized interaction patterns and refined requirements before being translated into a coherent click prototype. Finally, a qualitative user study, using the think-aloud method and semi-structured interviews based on the built prototype, assessed how individualized moderation influences users' perception and experience. Using the Website Accessibility Conformance Evaluation Methodology (WCAG-EM), the prototype was tested against WCAG 2.2 AA at design level. Eight design principles (DP) were identified to guide the creation of future middleware. Among these, transparency and explainability (DP8) emerged as a precondition for the other seven rather than a parallel requirement. A further pattern concerns the relationship between direct moderation configuration and features that operate at scale, such as moderation list subscription, which require considerably more user trust. The study also points out that generative artifacts, such as the discovery workshop and the wireframes, surface users' wants, but only an evaluative artifact, such as the interactive prototype or the qualitative study built on it, can surface what it costs users to act on those. While this thesis focuses on middleware, some findings are transferable to interactive recommender systems more broadly, depending on the scope and behavior of the algorithm and the amount of discovery content users are exposed to.