Abstract
Background The rapid adoption of Transformer-based architectures such as BERT and RoBERTa has advanced text classification in high-stakes domains, including cyberbullying detection, mental-health screening, and fake-news moderation. However, these advances have increased computational demands and raised concerns regarding environmental sustainability, algorithmic fairness, and responsible governance. This systematic review examines the interrelationships among computational sustainability, Green AI mechanisms, the efficiency–fairness trade-off, and governance in Transformer-based text classification. Methods A systematic literature review was conducted following the PRISMA 2020 guidelines. Studies were identified from five databases: ScienceDirect, Wiley Online Library, IEEE Xplore, Scopus, and ProQuest. A total of 94 studies published between 2020 and 2026 were synthesised to examine strategies for improving computational efficiency, their implications for fairness, and governance mechanisms relevant to responsible Transformer-based text classification. Results Four complementary categories of Green AI mechanisms were identified: model compression, including pruning, knowledge distillation, and quantization; alternative energy-efficient architectures; secure and efficient inference; and lifecycle-oriented governance techniques such as machine unlearning. Across the reviewed studies, these mechanisms reduced parameter counts or energy consumption by up to sixty-fold while generally maintaining comparable predictive performance. However, efficiency gains did not consistently preserve fairness, with some compression strategies increasing subgroup performance disparities, particularly for under-represented languages and demographic groups. Parameter-efficient fine-tuning demonstrated potential for jointly improving efficiency and equity. Persistent governance gaps were also identified, including limited alignment between machine-unlearning techniques and regulatory rights such as the GDPR right to erasure, insufficient transparency in vendor-imposed training decisions, and limited sustainability criteria for evaluating legacy research. Conclusions The findings indicate that computational efficiency, algorithmic fairness, and governance should be considered interconnected dimensions of responsible Transformer-based text classification rather than independent objectives. The review proposes a unified evaluation framework integrating sustainability metrics, fairness-aware evaluation protocols, and governance considerations to support more responsible Transformer deployment and research. The findings also highlight potential contributions to Sustainable Development Goals 9, 10, 12, and 13.