Cyberbullying on social media platforms remains a serious threat to digital well-being, requiring intelligent systems capable of detecting both explicit and subtle, emotionally charged abuse. Sentiment analysis (SA) plays a key role by interpreting emotional tone, polarity, and context, offering more nuanced and timely detection than keyword-based models. Emotions like anger, sarcasm, or veiled hostility often precede cyberbullying, especially during impulsive interactions. SA captures these affective cues, improving sensitivity to implicit abuse and coded language. This study presents the first systematic comparison of sentiment-enhanced transformer models such as ALBERT, DeBERTa, ELECTRA, HateBERT, and DeepSeek-coder-1.3b-base, fine-tuned for cyberbullying detection across Twitter (currently X), IMDB, and Amazon. Models were evaluated on predictive performance (Accuracy, Precision, Recall, F1-score), time and cost efficiency (inference time, memory, CPU/GPU use, and energy). ELECTRA + SA outperformed all models, achieving 91.85% accuracy, precision, and recall, and a 91.84% F1-score. It also excelled in efficiency, with 0.069 seconds inference time, 23.92 MB RAM use, 7.2% CPU/GPU usage, and 0.000075 kWh energy consumption, proving highly generalizable, sentiment-sensitive, and suitable for real-time, resource-aware deployment. These results highlight the importance of sentiment integration, dataset diversity, and computational efficiency in building scalable, real-world cyberbullying detection systems.

Cyberbullying significantly impacts mental health by adversely affecting victims’ psychological well-being. It is a prevalent issue on social media platforms, necessitating effective real-time detection systems to identify harmful content. However, current detection systems face challenges related to performance, dataset quality, time efficiency, and computational costs. This study compares existing text classification techniques for cyberbullying detection, evaluating their effectiveness on social media platforms. Large language models such as BERT, RoBERTa, XLNet, DistilBERT, and GPT-2.0 are assessed for their suitability. Results show that BERT achieves optimal performance, with 95% accuracy, precision, recall, and F1 score; a 5% error rate; 0.053 seconds inference time; 35.28 MB RAM usage; 0.4% CPU/GPU utilization; and 0.000263 kWh energy consumption. These findings highlight that while generative AI models are powerful, fine-tuned models often outperform them when adapted to specific datasets and tasks.

Sustainable AI considers both the environmental impact of AI technologies and their role in advancing global sustainability. This study provides a comprehensive analysis of emerging trends in sustainable AI, highlighting energy-efficient algorithms, green data centers, AI-driven resource management, and ethical governance. It emphasizes AI's dual nature, its potential to drive climate action and socio-economic progress, alongside risks such as carbon emissions, e-waste, and digital inequality. The study underscores the need for interdisciplinary collaboration, inclusive policies, and standardized sustainability metrics to ensure responsible AI deployment. By identifying barriers and proposing strategies, it offers guidance to researchers, policymakers, and industry leaders on aligning AI with long-term environmental, ethical, and social objectives.