This paper investigates how AI infrastructure inherently disadvantages speakers of underrepresented languages, using Bengali as a case study. It identifies four key structural failures—web presence, training token deficit, tokenization penalty, and connectivity exclusion—that lead to systematic exclusion and reduced access to AI-assisted educational tools.

This paper introduces a two-stage approach for adapting Qwen2.5-3B to Tibetan, a low-resource language, using Continual Pretraining (CPT) for linguistic grounding and Supervised Fine-Tuning (SFT) for task specialization. The study demonstrates significant improvements in perplexity and translation quality, along with an in-depth analysis of parameter evolution during adaptation.