Introduction
The rapid proliferation of digital media has made the detection and moderation of objectionable content a critical challenge for online platforms. This white paper, authored by a multi-disciplinary team from the University of Houston and other major institutions, outlines a strategic framework for addressing this issue. By synthesizing insights from computer science, psychology, and communication, the authors propose a nuanced taxonomy designed to move beyond simple keyword filters toward sophisticated, multimodal AI systems capable of understanding the "why" and "how" of online harm.
The Challenge of Modern Content Moderation
Current content moderation strategies rely heavily on two flawed pillars: human judgment and simplistic automated filters. Human moderation, while nuanced, is fundamentally unscalable given the massive volume of content uploaded to platforms like TikTok and YouTube. Conversely, existing automated filters often fail to identify "wolf in sheep’s clothing" content—where adult themes are masked using familiar children's characters, a phenomenon exemplified by the "Elsagate" controversy.
The authors argue that "objectionable content" lacks a rigorous, research-backed definition in the context of machine learning. Harm is often multimodal (a combination of specific audio, text, and imagery) and implicit (embedded in "meta-messages" rather than graphic visuals). This paper seeks to bridge the gap between psychological research on media effects and the technical requirements for building automated detection systems.
A Multidisciplinary Taxonomy
The core contribution of this work is a comprehensive taxonomy of objectionable content, which serves as a roadmap for future research and development. The authors identify five core categories of content with documented negative impacts, particularly on younger audiences: violence, sexual messages, the pairing of pleasure with violence, misguided meta-messages, and verbal/social aggression.
The taxonomy (Figure 1) provides a hierarchical model for classification. The authors propose a two-stage annotation strategy: first identifying explicit content (graphic violence, language), followed by the more complex task of identifying implicit content (embedded meta-messages).
Beyond Explicit Triggers: Implicit Harms and Meta-messages
A key insight of the paper is that severity is a continuum rather than a simple dichotomy. The authors highlight several nuanced forms of harm:
- The Danger of "Justified" Violence: "Realistic" war movies with dire consequences may be less harmful than superhero movies where violence is "justified" and lacks negative consequences for the protagonists, as the latter more easily encourages behavioral changes in viewers.
- Sarcasm as Aggression: Popular sitcoms can normalize the idea that "friendly" interaction involves being mean to others, which the authors link to the roots of toxic masculinity.
- Implicit Harms: Objectionable content isn't just about what is shown, but what is missing—such as the lack of health-risk depictions in sexual content or the "opportunity cost" of screen time leading to physical health issues.
Future Directions
While this white paper establishes a foundational framework, several challenges remain. Future work will focus on mapping the severity of content, which is a complex but crucial dimension. Additionally, detecting implicit meta-messages remains significantly more challenging than identifying explicit graphic content. The authors also aim to expand the system to detect "pro-social" content—media that inspires compassion and kindness—to provide a more holistic approach to content moderation.
This framework serves as the basis for a massive, multimodal video repository intended to support the development of next-generation AI moderation tools.
Summary updated