Social networks function as a means of communication, networking, information dissemination, and exchange. Unfortunately, they are also a misused space for the spread of hate speech and cyberbullying, making them a delicate global issue as they facilitate attacks on the integrity and dignity of individuals, particularly women, adolescents, and girls. This study explores methods, techniques, and recent advances in artificial intelligence models developed to prevent, detect, and mitigate hate speech on online social networks. To achieve this, we systematically reviewed the literature using the PRISMA methodological framework, which provides a structured and methodical approach. The review focused on articles using machine learning (ML) techniques and Natural Language Processing (NLP) algorithms, resulting in the selection of 33 primary articles. The findings show that the BERT model was the most widely used in this field. However, the models with the highest precision are FAST-RNN and BREE-HD, with precision indicators of 98% and 97%, respectively. This research identified new avenues for multilingual models for hate speech detection and the generation of synthetic data to train these models.