Hate speech, which disparages individuals or groups based on identity traits such as race, ethnicity, or religion, poses significant societal risks. This paper presents a machine learning approach to detecting Arabic hate speech on social media using a balanced dataset of 8,000 comments categorized into regular text, insults, pejoratives, and cyberbullying. A Textual Labeled Polarity Algorithm, developed with dedicated hate speech detection lexicons and incorporating authorial feedback, was employed. Three machine learning algorithms Naive Bayes (NB), Support Vector Machines (SVM), and K-Nearest Neighbors (K-NN) were tested. SVM was the most effective, achieving an accuracy of 92%, while NB and K-NN reached accuracies of 82.5% and 81.5%, respectively. The effectiveness of these models demonstrates their potential for real-world applications in social networks as AI protection tools, ensuring safer browsing experiences by detecting and hiding hate speech. This research underscores the potential of AI in enhancing digital security and content moderation practices, suggesting a scalable approach for social media platforms.