This systematic review analyzes the effectiveness of deep learning models in detecting phishing attacks on social media based on 33 articles selected from the Scopus database. The main problem identified was information and credential theft, which was present in 84.8% of the studies, followed by reputational damage, malicious content distribution, and financial impact. The proposed objectives are to mitigate phishing, improve model accuracy and efficiency, and develop deep learning-based applications. Hybrid models were the most widely used (32.4%), followed by RNN, CNN, Transformers, Autoencoders, and GAN. The most used evaluation metrics were accuracy (29.4%), precision (25.5%), recall (23.5%), and F1 score (21.6%), whereas classic dataset splitting (26 articles), the use of public datasets (21), and K-fold cross-validation (9) were the most frequently used methods. The reported performance levels show promising results, with accuracy between 90% and 99%, F1 scores above 85%, and better performance in hybrid models and architectures such as BERT. However, the review highlights important limitations, such as the lack of generalization to real-world scenarios, dependence on specific datasets, and limited evaluation of multimedia content. The complexity of the digital social environment, where high variability, dynamic human behavior, and constant interaction limit the effectiveness of traditional techniques, was identified as a key challenge. The research concludes with the need to encourage longitudinal studies, expand the use of heterogeneous data, and promote practical solutions that strengthen cybersecurity in social media through adaptable and robust deep learning models, particularly recommending the use of hybrid approaches and federated learning techniques as promising avenues for future research and application.