Multilingualism is a design approach to artificial intelligence (AI) that safeguards the pluralism of knowledge in the digital sphere. Emerging language technologies, in particular large language models (LLMs), construct a spatial map of meaning in their neural networks using word-embedding techniques. These encodings represent the relationship between words and concepts through numerical values of distance and direction. LLM-generated text reflects the conceptual priorities of the language it was trained on, meaning multilingualism is a structural question when LLMs influence data flows. The 2003 United Nations Educational Scientific and Cultural Organization recommendations on access to cyberspace declare digital multilingualism as essential to information rights. Upholding these norms today requires that linguistic diversity, and the underlying geometries of knowledge held in language, shape the AI outputs that contribute to contents of the internet.
