This study presents a systematic review of the literature (SLR) on the methodologies and techniques used to detect cyberbullying on social networks to identify the most effective approaches and the main challenges in the area. The analysis of recent studies found that approximately 60% of the research has used natural language processing (NLP) algorithms, while 40% have opted for machine learning (ML) techniques for automatic detection. The review also revealed that 50% of the studies identified limitations related to the quality and representativeness of the available data sets. At the same time, 30% highlighted the lack of integration of more advanced approaches into existing models. Regarding the challenges, we observed that 70% of the works pointed to the need to improve the precision in detecting cyberbullying and reduce biases in the models. In addition, this study recommended the development of hybrid approaches that combine supervised and unsupervised learning, which could enhance the effectiveness of the models. Furthermore, the results indicate that the collaboration between the research community and the social media stakeholders should be strengthened, which could reduce classification errors and open up new research lines for more accurate and effective future solutions.