AI Infrastructure Fails Bengali Speakers: Structural Barriers in Language Tools

A recent study published on arXiv (2608.12278) highlights the systematic disadvantages faced by speakers of less-represented languages in AI infrastructure, using Bengali as a focal point. Although Bengali is spoken by nearly 4% of the world’s population, it constitutes less than 0.5% of online content. The research uncovers four interconnected failures: a significant gap in web presence, a 67:1 disparity in training tokens between English and Bengali in key multilingual datasets, challenges related to Bengali's alphasyllabary script, and obstacles in deployment for areas with low connectivity. These structural challenges arise even before model training, complicating the potential of AI to enhance education and language support in underserved communities. The paper emphasizes the importance of inclusive AI practices that prioritize linguistic diversity.