From Black-Box Filters to Agentic Pipelines: Designing Calibrated Reliance in Video Moderation

Generative AI has turned online video into a high-stakes environment, triggering a collapse in digital trust. To manage User-Generated Content, publishers are adopting "agentic pipelines"—multi-step AI systems autonomously invoking tools to execute cascading moderation decisions. However, their opacity introduces severe accountability risks. In this position paper, we argue the HCXAI community must shift toward designing for calibrated reliance within "Trust-to-Action Moments". Explainability must serve two stakeholders: editorial moderators overseeing auto-publishing, and end-users needing transparent provenance for relational engagement. Drawing on the Vialog moderAId pipeline feasibility study, expert publisher interviews (N=8), and an experimental study on Psychological Ownership (N=499), we propose three sociotechnical requirements for agentic explainability: explainable traceability, configurable sensitivity, and progressive delegation. We provoke the community to move beyond single models, designing instead friction-tuned verification flows as accountability infrastructure for digital discourse.