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

Paper Title

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

Publication Info

• Topic area: Personalized content moderation and user safety in online platforms.
• Keywords: content moderation, personalized transformation, user agency, harm reduction, online safety, phobias, trauma, user-centered design, AI-driven intervention, exposure therapy.

Background and Problem

• Problem / challenge: Current content moderation approaches are centralized and suppressive, failing to address the subjective and contextual nature of harm. They force users to choose between enduring distress or disconnecting entirely, often removing valuable content alongside harmful elements.
• Significance: This issue impacts users’ mental health, agency, and ability to engage with online communities, especially for those with specific sensitivities like phobias, PTSD, or trauma-related triggers.
• Motivation and related work: Prior work has focused on centralized moderation, subjective harm, and user-facing tools, but these approaches still rely on suppressive mechanisms. There is a gap in enabling users to transform harmful content while preserving its informational value, which this paper addresses.

Solution

• Proposed approach: DIY-MOD (Do-It-Yourself Moderation), a browser extension that enables personalized content transformation by modifying sensitive elements in real-time based on user-defined filters.
• Novelty:
  1. Introduction of a new paradigm: personalized content transformation that preserves informational value while reducing harm.
  2. Design and implementation of DIY-MOD, a system that allows users to define and apply nuanced filters to transform content.
  3. Development of a multi-modal intervention palette for text and images, offering transformations like obfuscation, inpainting, and artistic stylization.
  4. Evaluation through two user studies, demonstrating increased user agency, safety, and engagement.
• Procedure and key techniques:
  • Users create filters via a conversational interface, specifying sensitivities in natural language or through examples.
  • The system intercepts and analyzes content using large vision-language models (LVLMs) and applies transformations like blurring, inpainting, or stylistic alterations.
  • A two-stage pipeline selects the most appropriate intervention based on semantic fidelity, trigger fidelity, perceptual smoothness, and contextual harm risk.
  • Transparency indicators mark modified content, and users can override transformations.

Results

• Concrete findings:
  • DIY-MOD increased users’ sense of agency and safety during naturalistic browsing.
  • 63% of image transformations and 79.1% of text transformations matched user preferences in controlled evaluations.
  • Manual audits showed a 93% success rate in correctly transformed images, with a 7% false-positive rate.
• Advantage over baselines:
  • Unlike traditional moderation, DIY-MOD preserves valuable content while reducing harm, enabling users to engage with content they would otherwise avoid.
  • Provides granular control and personalization, addressing the limitations of centralized, one-size-fits-all moderation.
• Experiments / evaluation:
  • Study 1: In-situ evaluation with 15 participants browsing Reddit, demonstrating increased control and reduced distress.
  • Study 2: Controlled preference elicitation with 12 participants, identifying principles like cognitive closure and contextual appropriateness for effective transformations.
• Limitations and future work:
  • Limited participant scale and cultural diversity.
  • Technical dependencies on platform-specific adapters and commercial LVLMs.
  • Scalability challenges due to computational costs.
  • Future directions include improving intervention selection models, exploring long-term psychological impacts, and refining privacy and collaboration features.

Summary

This paper introduces DIY-MOD, a browser extension that operationalizes personalized content transformation to address the limitations of centralized, suppressive moderation. By allowing users to define and apply nuanced filters, DIY-MOD transforms harmful content elements while preserving their informational value. Two user studies demonstrate that this approach increases user agency, safety, and engagement with online communities. While challenges remain in scalability, cultural generalizability, and long-term evaluation, DIY-MOD represents a significant step toward user-centered, adaptive moderation tools that prioritize individual well-being.