Diversity Diminishes with Language Disparity

Low-resource languages, like Tigrinya and Bengali, face challenges in training Large Language Models (LLMs) due to the lack of high-quality data available. The variance in resource allocation leads to superior performance for languages like English, while others suffer noticeably, resulting in a bias in a model's responses. This uneven performance may cause users to inadvertently adopt a limited perspective aligned with English speakers, challenging the richness of diverse thought. AI researchers are concerned that this standardization of English usage in multimodal tasks could significantly hinder scientific diversity and creativity.