False narratives are a serious threat to the reliability of information and public trust, especially when widely spread through mass media. The paper proposes a deep learning framework to classify fake and real news articles based on examining four consecutive neural network architectures, such as LSTM, GRU, Bidirectional LSTM, and Bidirectional GRU. The models were trained by using a large-scale training set of 44,898 news articles, the textual data in the model was preprocessed, which contained lowercasing, noise cleaning, stopword elimination, and sequence padding. The Bidirectional GRU performed best and outperformed the other architectures by having an almost perfect score of 98.38% and was able to capture untidily well the context dependency in both directions. The model included insertion of embedding layers, dropout regularization, and global max pooling in learning efficiency and minimizing overfitting. The extensive analysis in terms of accuracy, precision, recall, F1-score, confusion matrix analysis, and plotting of lexical patterns proved the efficiency of the model. The results confirm the viability of using such architectures to effectively and efficiently detect fake news as experienced in the media.