New AI Detects Pure Synthesis Fake News Videos

This paper addresses the growing challenge posed by advanced text-to-video (T2V) generation models, which can now create entirely synthetic fake news videos. Unlike "cheap fakes" that repurpose existing footage, these new "pure synthesis" fakes are generated from scratch, making them highly aligned with fabricated narratives and difficult for current detection systems to identify. The authors highlight that existing datasets lack these pure synthesis videos, and simply prompting T2V models for fake news can lead to detection shortcuts and visual quality issues. To counter this, the research introduces a new ternary classification task for fake news video detection (FNVD), categorizing videos as real, cheap fake, or pure synthesis fake. They also developed the first pure synthesis fake news video dataset (PS-FNVD), which includes both fabricated events with aligned deception and true events with false visual provenance, preventing models from relying on unimodal shortcuts. Furthermore, they propose the Reasoning-guided T2V-FNVD (R-T2V) framework. This framework integrates high-level semantic logic with low-level generative traces, trained through conditioned rationale generation and supervised fine-tuning, to predict the veracity label. Extensive experiments demonstrate R-T2V's superior performance, significantly outperforming ten prevailing baselines in accuracy and macro F1 score. This advancement provides a crucial tool in the fight against sophisticated AI-generated disinformation.