An Actionable Strategy for Assessing Multilingual Drift and Fairness in Low-Resource Languages

Large language models are increasingly applied across diverselanguages, yet their behavior in low-resource settingsremains insufficiently examined. Semantic drift, shifts in sentiment,and loss of culturally grounded meaning often occurwhen models process underrepresented languages, with consequencesthat extend beyond technical error. Existing evaluationtools require substantial datasets or compute, placingthem out of reach for many Global South researchers. Thispaper presents a lightweight, accessible strategy for assessingmultilingual drift and fairness using small paired examples,simple embedding-based checks, and culturally informed criteria.Rather than proposing new metrics, the approach offersa practical template that can be adapted to local needs anddeployed with minimal resources, enabling more equitablescrutiny of model behavior across languages.