Explainable multilingual and multimodal fake-news detection: toward robust and trustworthy AI for combating misinformation

Fake-news detection requires systems that are multilingual, multimodal, and explainable-yet the majority of the existing models are English-centric, text-only, and opaque. This study introduces two key innovations: (i) a new multilingual-multimodal dataset of 74,000 news articles in Hindi, Gujarati, Marathi, Telugu, and English with paired images, and (ii) Hybrid Explainable Multimodal Transformer Fake (HEMT-Fake) that integrates text, image, and relational signals with hierarchical explainability. The architecture combines transformer embeddings, a convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) text encoder, residual network (ResNet) image features, and graph sample and aggregate (GraphSAGE) metadata, all of which are fused via multi-head attention. Its explainability module unites attention, Shapley Additive exPlanations (SHAP), and local interpretable model-agnostic explanations (LIME) to provide token-, sentence-, and modality-level transparency. Across four languages, HEMT-Fake delivers a ~ 5% Macro-F1 improvement over Cross-Lingual Language Model with RoBERTa (XLM-R) architecture and Multilingual Bidirectional Encoder Representations From Transformers (mBERT), with gains of 7-8% in low-resource languages. The model achieves 85% accuracy under adversarial paraphrasing and 80% on artificial intelligence (AI)-generated fake news, halving robustness losses compared to baselines. Human evaluation reveals that 82% of explanations are judged to be meaningful, confirming transparency and trust for fact-checkers.