SMS spam continues to pose significant security and usability challenges, especially in regions where text- based communication remains dominant. Traditional spam filtering approaches often rely on surface-level lexical patterns that fail to capture contextual semantics. This study explains a comparative framework that demonstrates multiple classifiers trained on a unified RoBERTa embedding representations. Results demonstrate that RoBERTa-based features significantly improve classification performance. Beyond technical performance, this study also situates SMS spam detection within its broader societal context, examining how spam disproportionately harms vulnerable populations, undermines digital trust, and impedes equitable access to mobile services in developing regions. The study highlights RoBERTa's strength in capturing contextual meaning, reducing manual feature engineering, and enabling scalable spam detection.