Beyond Accuracy: A Hybrid BERT-BiLSTM Framework with Explainable AI (XAI) for Detecting Machine-Generated Disinformation

The rapid rise of Large Language Models (LLMs) has shifted the battleground of digital misinformation. Unlike human-written fake news, machine-generated disinformation often employs subtle linguistic patterns that evade conventional detection systems. Although Deep Learning models can effectively identify synthetic text, they frequently operate as "black boxes," failing to offer the transparency needed for sensitive real-world applications. To address this, we introduce a hybrid architecture that merges the contextual strengths of DistilBERT with the sequential analysis capabilities of Bidirectional Long Short-Term Memory (BiLSTM) networks. Crucially, we incorporate SHapley Additive exPlanations (SHAP) to decode the model's decision-making process, visualizing exactly which words or tokens tip the scales toward a specific classification. Tests on the benchmark Fake or Real News dataset [1], supplemented by a 5-fold cross-validation protocol to ensure robust statistical validation, show our framework achieves an average accuracy of 96.92% ± 0.18%. By leveraging Explainable AI (XAI), we confirm that the model identifies actual semantic anomalies rather than merely overfitting to background noise, offering a more trustworthy foundation for automated fact-checking systems.