Uncovering inequalities in new knowledge learning by large language models across different languages

Significance Large language models (LLMs) are transforming daily life, yet users across different languages may not benefit equally. Our study shows that LLMs face greater challenges in learning new knowledge and resisting incorrect or misleading information in lower-resource languages. Consequently, users of these languages are at higher risk of receiving lower-quality or misleading outputs. Moreover, knowledge is more easily transferred to, and more likely to be retained and prioritized in, higher-resource languages. As AI systems become increasingly integrated into society, such disparities threaten to marginalize lower-resource languages and deepen global information inequality. We also analyze their underlying causes and propose preliminary strategies for mitigation. Addressing these inequalities is essential to build AI systems that are fair, inclusive, and socially responsible.