What If Moderation Didn't Mean Suppression? A Case for Personalized Content Transformation

TL;DR

The paper argues that centralized content moderation fails to account for subjective harm and often suppresses valuable content. It introduces DIY-MOD, a browser extension that enables personalized content transformation by selectively modifying distressing elements while preserving meaning, guided by user-defined filters and a two-stage AI-assisted selection pipeline. Through two studies, the authors show increased user agency, safety, and engagement, and extract design principles (e.g., cognitive closure, transparency, and context-aware transformations) for effective interventions. The work advocates platform-integrated personalization as a scalable path while emphasizing privacy-by-design, open collaboration, and careful consideration of authorship and civic discourse implications. Overall, it offers a practical middleware approach that could reshape how individuals navigate harmful content without full disengagement from online communities.

Abstract

Centralized content moderation paradigm both falls short and over-reaches: 1) it fails to account for the subjective nature of harm, and 2) it acts with blunt suppression in response to content deemed harmful, even when such content can be salvaged. We first investigate this through formative interviews, documenting how seemingly benign content becomes harmful due to individual life experiences. Based on these insights, we developed DIY-MOD, a browser extension that operationalizes a new paradigm: personalized content transformation. Operating on a user's own definition of harm, DIY-MOD transforms sensitive elements within content in real-time instead of suppressing the content itself. The system selects the most appropriate transformation for a piece of content from a diverse palette--from obfuscation to artistic stylizing--to match the user's specific needs while preserving the content's informational value. Our two-session user study demonstrates that this approach increases users' sense of agency and safety, enabling them to engage with content and communities they previously needed to avoid.

Figures (5)

• Figure 1: Intervention palette demonstrating DIY-MOD's three categories of transformation, applied to an original image (a). Semantic Modification[b, c] alters the content through techniques like (b) inpainting, which removes the trigger and (c) replacement with user-specified alternatives---trees, stars, cards in this case. Obfuscation[d] reduces trigger fidelity, shown here with blurring (d). Stylistic Alteration[e, f] changes the rendering style; a Studio Ghibli style animation is shown applied to the entire image (e) and selectively to only the trigger region (f). The transformations shown address the filter created by Rex (\ref{['sec:intro']}), who is managing binge-eating disorder.
• Figure 2: DIY-MOD's two-stage intervention selection pipeline. Content matching user filters enters Stage 1 (pruning) to identify promising interventions, then Stage 2 generates and scores K candidates before selecting the best transformation. Non-matching content bypasses the pipeline entirely (paths not shown for clarity).
• Figure 3: The custom application used in Study 2 presented participants with a side-by-side, synchronized dual feed. Each post appeared as a pair of distinct modifications. Participants scrolled the feed and, for each post, used the controls at the bottom to indicate their preference for the text and image transformations separately. In this example, the user has preferred the text transformation on the left (a full overlay) but preferred the image transformation on the right (occlusion). Modification indicators were not shown in this interface, as participants were informed that all content was transformed by the system.
• Figure 4: DIY-MOD processing pipeline showing synchronous and asynchronous execution paths. The critical path (diagonal hatching) completes in 5-15 seconds for a batch of posts, returning text modifications and polling tokens. Image transformations process asynchronously in background workers (dotted pattern). The client polls for completed transformations—missing on attempts until the content transformation is ready. Previously processed content can return immediately from cache on first poll. Note that, the critical path delay is for a batch of usually 25 posts in reddit. And, this delay only happens when the user first loads/refreshes the feed in browser. For subsequent batches of posts within the browsing session, predictive prefetching (\ref{['appendix:prefetching']}) minimizes the delay under reasonable scrolling behavior.
• Figure 5: DIY-MOD's filter creation and configuration interface (\ref{['fig:chat-interface']})Chat interface for filter creation.DIY-MOD engages in conversational grounding to help users specify their content sensitivities. When a user enters a broad term, the system can prompt for clarification about more specific aspects. Users control specification depth through suggested options or free text input. This chat interface opens as a popup when user clicks on the extension icon. (\ref{['fig:filter-config']})Filter configuration interface. After establishing the filter description, users configure three parameters: (1) Content Type specifies whether the filter applies to text, images, or both; (2) Sensitivity Level indicates the user's distress intensity; and (3) Duration sets filter expiration. For brevity, in popup interface we only showed three time presets. Users can modify the filter description anytime using the "Modify Filter" button or add metadata through the Options page for further customization.