A Multilingual Paradigm for AI Created by All, for All

In a first of its kind, the 2026 New Delhi Frontier Model Voluntary Commitments launched at the India AI Impact Summit featured a clear, albeit nonbinding, requirement that AI model providers conduct multilingual evaluations of models. Establishing a commitment about multilingual performance and evaluations was no easy feat and was a testament to years of research and advocacy by natural language processing (NLP) experts, civil society advocates, and academics. It was also a massive step forward: a recognition that the current paradigm of AI testing, particularly AI model evaluations, doesn’t sufficiently capture the nuances of languages spoken and contexts across the world.

Currently, AI models, specifically large language models (LLMs), do not work equally well across the world’s ~7,000 languages. According to the International AI Safety Report, advancements in AI technologies are “jagged,” and especially uneven when it comes to how poorly systems perform in languages other than English. As AI systems—particularly LLMs powering information systems like chatbots, content analysis tools, and even decision modeling tools—become increasingly integrated into the social fabric by governments, healthcare institutions, academic centers, industry, and more, the need to ensure that these systems work equally well across languages is essential. Yet, the conversations dedicated to the development, governance, and use of these LLM-based systems are largely contained to the Global North and state or corporate-led spaces. This results not only in poorer tools, but fewer opportunities to shape the tools for the Global Majority with some experts ringing the alarm of a possibly widened “digital language gap.”