The widespread use of social media and online news platforms has made it easier for misinformation and fake news to spread rapidly. This creates serious challenges for individuals and organizations that rely on accurate information. To address this problem, this study proposes a fake news detection system that combines Natural Language Processing (NLP) techniques with both traditional machine learning and transformer-based models. The dataset used for the study is derived from the WELFake dataset, containing labeled news articles categorized as real or fake. Text preprocessing techniques such as tokenization, removal of noise, and normalization are applied to prepare the data. Traditional models like Support Vector Machine (SVM) and LightGBM use TF-IDF features to capture important word patterns, while DistilBERT is used to understand contextual meaning in text. The results show that transformer-based models achieve higher accuracy, while traditional models remain efficient and reliable. This hybrid approach improves the overall effectiveness of fake news detection systems.