Abstract
With the increasing prevalence of deepfake content across social media, there is a growing challenge to trust online-especially when it comes to analyzing user attitude towards manipulative pieces of text. However, with millions of tweets generated every day, separating genuine sentiments from artificial narratives is a challenging area in sentiment analysis research. While previous models are often highly accurate, they depend on feature engineering and act like a black-box system, compromising interpretability and generalizability. This underscores the importance of having frameworks that are robust but also explainable. In the study, we evaluate the proposed model using TweepFake dataset which is publicly available with human and AI-generated tweets. Our study adopted the TweepFake dataset, which contains 25,572 tweets equally distributed between human and AI generated. Its balanced structure guarantees that the different methods of deepfake sentiment detection are assessed fairly. We present a hybrid method, which combines semantic richness (captured through Transformer-based contextual embeddings) with human-readable explanations based on fuzzy membership rules embedded in a Fuzzy Rule-Based System (FRBS). Its novelty is in integrating state-of-the-art contextual modeling with symbolic reasoning for interpretability. Under 10-fold cross-validation, performance was evaluated through Accuracy, Precision, Recall, F1-score, Jaccard Coefficient & MCC and Inference time. Experimental results demonstrate that the new framework has 97.0% accuracy and F1-score, with stable performance over epochs and more explainability compared to baseline models. The model outperforms the RoBERTa baseline, achieving a 1.6% improvement in accuracy.