In this study, we created and tested a sophisticated fake news detection system using a proposed model called VeritasBERT, which combines Bidirectional Encoder Representations from Transformers (BERT) with a softmax classification layer. Our methodology included extensive data collection using web crawlers, preprocessing, and feature extraction in order to effectively train the model on a diverse dataset derived from various online sources, such as social media platforms and news sites. In comparison to traditional and existing models, the proposed model outperformed them on a variety of metrics. The VeritasBERT achieved 99.49% accuracy, 99.38% precision, 99.24% recall, and an F1 score of 99.45%. These results outperformed other tested models, including LSTM, NLP combined with LSTM, and hybrid models like Naive Bayes with Logistic Regression, with the highest accuracy of 99.2%, precision of 99.1%, and recall of 99%. Precision-recall curves demonstrated the model’s efficiency by confirming its high precision even at varying recall thresholds, making it extremely reliable for practical applications. Error analysis revealed that the model’s main challenges are distinguishing between ‘Suspicious’ and ‘Real’ news, which will guide future system improvements. This study demonstrates the effectiveness of using advanced deep learning techniques to address the complexities of fake news detection, providing robust solutions that can be applied to real-time news verification systems to improve information credibility across digital platforms.