A Novel Responsible AI Model for Fusion Fake News Detection

The rapid proliferation of fake news on social media threatens public trust and social stability. The authors propose a Responsible AI (RAI) framework for multimodal fake news detection that integrates threat theory with deep learning. They introduce two responsibility indicators: (1) User Trust Score (UTS), measuring source credibility via trustworthiness, expertise, and behavioral consistency; (2) Responsible News Dissemination Score (RNDS), quantifying content alignment with responsible discourse using a 66-keyword RAI vocabulary. These indicators are integrated into Graph Neural Networks through node-level feature augmentation and responsibility-aware message passing. The model fuses BERT text, VGG19 visual features, propagation patterns, and responsibility metrics, achieving F1-scores of 0.9498 on Politifact and 0.9720 on Gossipcop. They discuss ethical risks including algorithmic stigmatization, privacy concerns, and demographic bias, and propose mitigation strategies such as temporal score decay, federated learning, and fairness audits to ensure responsible deployment.