<p>The rapid growth of online media has accelerated the circulation of misinformation, posing serious risks to public trust, social stability, and democratic processes. Detecting fake news remains a challenging task, particularly because traditional machine learning approaches such as Support Vector Machines and Naïve Bayes rely on shallow linguistic cues and often fail to recognize sarcasm, implicit tone, or evolving writing styles. Earlier deep learning models, including CNNs and LSTMs, provide modest improvements but continue to struggle with long text and shifting topical contexts. This study examines the effectiveness of two transformer-based architectures—BERT and DeBERTa—for identifying fake news using only textual information from headlines and article bodies. To enhance generalization, news articles from ten publicly available datasets are consolidated into a single corpus namely FakeDiverse, that exposes the models to a broader spectrum of linguistic patterns. MinHash similarity along with Locality Sensitive Hashing algorithm has been employed to evaluate the lexical redundancy within the FakeDiverse corpus as a post hoc analysis. Class imbalance between real and fake news is addressed through a weighted cross-entropy loss function, ensuring that minority-class instances are adequately represented during training. FakeDiverse dataset is split in the ratio of 80:20 for training and testing, and subsequently the models BERT and DeBERTA are trained on the training set and evaluated using the test sets.&#xa0;Experimental results demonstrate that both the models capture contextual nuances effectively, with BERT achieving 98% accuracy and DeBERTa reaching 99%. The findings demonstrate that BERT and DeBERTa achieves strong performance on FakeDiverse corpus, and cross-dataset evaluation shows that models have challenges to generalize unseen data distribution thereby highlighting the need for enhanced generalization strategies as well as domain adaptation.</p>