Multimodal Artificial Intelligence for Social Media Content Verification, Influence Assessment and Risk Analysis

Abstract

Social platforms now carry news, opinion, commerce and public safety information in a blend of text, pictures, video and audio. Deception has adapted to this format: a genuine photograph can be given a false caption, a synthetic voice can be attached to an old clip, and coordinated accounts can push a post until it looks popular. Detectors that examine only one channel often miss these cases. This paper presents a unified multimodal framework that treats three linked questions together: is the content authentic (verification), who and what is amplifying it (influence), and how much harm could it cause (risk). The framework encodes each modality separately, aligns them through cross-modal attention, and feeds a shared representation to three task heads whose outputs are merged into an interpretable risk score with a human review step. We survey existing work, describe the architecture, datasets, training approach and evaluation protocol, discuss applications in journalism, public health, brand safety, elections and disaster response, and examine ethical constraints such as privacy, bias and over-blocking. The contribution is a design and evaluation blueprint rather than a new benchmark result, and we close with open problems and a research agenda.