This paper presents CyberShieldAI, a novel framework that leverages BERT (Bidirectional Encoder Representations from Transformers) models to detect cyberbullying in online text communications. Our approach harnesses BERT’s contextual understanding capabilities to identify subtle linguistic patterns, emotional undertones, and semantic structures commonly associated with cyberbullying content. By fine-tuning BERT on carefully curated datasets of online interactions, CyberShieldAI can effectively distinguish between harmless communications and those containing bullying, harassment, or other harmful content. The framework analyzes text bidirectionally, capturing crucial contextual relationships that traditional text classification methods might overlook. Experimental evaluations demonstrate that our BERT-based system achieves significant improvements in cyberbullying detection accuracy, sensitivity, and specificity compared to conventional machine learning approaches. This paper details the system architecture, implementation methodology, dataset preparation challenges, and performance metrics that validate the effectiveness of our strategy. CyberShieldAI represents a significant advancement in automated content moderation systems, which can help create safer online environments, particularly for vulnerable user populations, such as children and adolescents.