The rise of fake news is a serious threat to public conversation and trust in society. This research tackles the issue of detecting automated fake news by creating and testing advanced deep learning models that classify news as either Fake or True across three different datasets: Fakeddit, LIAR 2, and Fake News Classification. The study uses a Multi-Task Learning(MTL) approach with RoBERTa for analysing test data, particularly focusing on the LIAR 2 dataset, and takes advantages of rich metadata to boost performance. For detecting fake news in a multimodal context on the Fakeddit dataset, a hybrid BERT+ResNet model is used to combine both text and visual information. Additionally, various machine learning and deep learning models are evaluated on the purely textual Fake News Classification dataset. The results show that RoBERTa(MTL) performs exceptionally well on LIAR 2, highlighting the benefits of using additional information. Meanwhile, the BERT+ResNet model offers a solid framework for handling multimodal data, but its performance on Fakeddit reveals the challenges of merging different types of information and spotting inconsistencies between text and images. On the other hand, the Fake News Classification dataset, which has simpler binary labels, shows high performance across a variety of models. This comparative analysis emphasizes how the characteristics and complexity of datasets significantly influence model effectiveness, offering valuable insights for developing future fake news detection systems.