This study develops a hybrid model that merges textual content and social context analysis to improve the accuracy of fake news detection, specifically in political and entertainment domains. The model combines a BERT-based textual analysis module with a Random Forest-based social analysis module. The textual module processes news content using tokenization and contextual embedding. In contrast, the social module assesses propagation patterns like retweets and likes alongside user metadata such as verification status and follower counts. A weighted voting system favors social analysis with a 55% weight, integrating the predictions from both components. Trained on the FakeNewsNet dataset, which includes news labeled by PolitiFact and GossipCop, the hybrid model achieved an F1 score of 86.4% on GossipCop and 80.7% on PolitiFact. This performance surpassed individual module scores (textual:  70%; social:  79–86%) and outperformed other methods like SVM-RNN and BerConvoNet. The results indicate that social analysis, particularly relevant for entertainment news, is more resilient to linguistic manipulation, yielding a 3–15% improvement over baseline approaches. The research underscores the importance of social propagation patterns and source credibility in combating misinformation. The proposed framework aims for scalable real-time detection that can influence platform moderation and policy. Future work will focus on enhancing adaptability and integrating multimodal data for better accuracy.