Misinformation becomes publicly available on social media and websites. Some journalists write fake news in different languages to target different societies. This study aims to develop an optimized algorithm for detecting and classifying fake news in the Arabic language. The proposed work investigates different deep learning techniques to identify effective and efficient algorithms that can handle fake news problems. The utilized techniques involve individual algorithms such as convolutional neural networks (CNNs), gate recurrent neural networks (GRUs), and bidirectional long short-term memory (BiLSTM). This work also considers hybrid algorithms such as CNNs with BiLSTM (CNN-BiLSTM) and CNNs with BiGRU (CNN-BiGRU). This research focused on Arabic news, motivated by the limited works interested in the Arabic language because of its linguistic difficulty. English has been widely investigated due to its simplicity. The proposed model used the Arabic Fake News Dataset (AFND) as an input channel for training and testing purposes. The model was trained to discriminate fake from real news. It achieved premium results with accuracies of 88.63%, 89.7%, 91.73%, and 92% for BiLSTM, BiGRU, CNN-BiLSTM, and CNN-BiGRU, respectively. The significant value achieved by CNN-BiGRU algorithms refers to the power of CNNs in feature extraction, the efficiency of the GRU structure and its simplicity in mitigating the vanishing gradient problem. This research aims to enhance model accuracy for unseen data in Arabic fake news detection.