Examining Potential Harms of Large Language Models (LLMs) in Africa

Abstract

Large language models (LLMs) have the potential to generate significant benefits, but their blanket application in Africa could exacerbate existing social and economic inequalities. This is due to a number of factors, including limited technological advancement, historical injustice and marginalization, and underrepresentation of African languages, values, and norms in training data. Despite comprising nearly one-third of the world’s languages, most African languages are underrepresented on the internet: they are primarily oral with little available in written and digitized form. Additionally, most African languages have conflicting orthographic standards. While Africa is undergoing a digital transformation, both internet connectivity and digital literacy remain relatively low and unevenly distributed. This lack of online representation for African languages limits the availability of natural language data for training inclusive language models. This paper examines the potential harms of LLMs in Africa, covering harms already documented for the African context; harms studied and documented for the Western context, but previously unapplied to Africa; and novel potential harms based on the norms, values, practices, and contextual factors of the African continent. This work aims to contribute to a better understanding of potential harms of LLMs in Africa, which in turn could inform and support the development of more inclusive LLMs.