Savoldi et al. (2025): A decade of gender bias in machine translation
Abstract A decade has passed since the first recognition of gender bias in machine translation in a seminal paper by Prof. Londa Schiebinger. Today, multilingual language technology, especially with the rise of advanced AI assistants powered by large language models (LLMs), increasingly shapes global communication and user interactions. Ensuring such technology is inclusive and does not result in harm has become critical, as gender bias in machine translation can lead to unequal representation, misgendering, and tangible service disparities that disproportionately affect marginalized groups. In this paper, we take stock of the last decade of research on gender bias in machine translation. We find that early optimism about a quick technological solution has given way to a more nuanced picture. Promising trends have emerged—the number of research efforts is growing, and the recognition of non-binary gender identities has improved—but significant challenges persist. These include an overemphasis on English-centric approaches and a tendency to disconnect translation technologies from the context in which they operate. While these efforts have made contributions, we argue that bias is dynamic, multifaceted, and resistant to simple solutions. We build upon the lessons of the past decade to discuss the current landscape in which LLMs are emerging. Our aim is to inspire future work in the field that transcends current limitations.Click here to access the article.
Citation Beatrice Savoldi, Jasmijn Bastings, Luisa Bentivogli, and Eva Vanmassenhove. 2025b. A decade of gender bias in machine translation. Patterns, 6.