BEYOND TRANSLATION: BUILDING HINDI-CENTRIC INTELLIGENCE FOR HUMAN–AI COLLABORATION

Abstract

Artificial Intelligence (AI) has become a powerful tool for communication and collaboration across languages. However, most AI systems remain English-centric, treating non-English languages as secondary through translation. For Hindi, one of the most widely spoken languages globally, this approach often results in the loss of cultural meaning, idiomatic richness, and socio-pragmatic depth. Translation tools can provide basic communication, but they fail to capture the nuances of Hindi expressions, dialects, and cultural references, leading to interactions that feel artificial or incomplete. This paper argues for a paradigm shift toward building Hindi-centric intelligence that treats Hindi as a primary language of thought rather than a peripheral translation target. We explore frameworks for embedding Hindi semantics, pragmatics, and socio-cultural cues directly into AI models. Through surveys of 200 participants across academia, industry, and government, we identify strong dissatisfaction with translation-based systems and a clear demand for authentic Hindi-centric AI. Experimental comparisons between translation-based tools and Hindi-trained models reveal significant improvements in accuracy, idiomatic interpretation, and user satisfaction. A key case study is India’s sovereign AI initiative, Sarvam LLM, developed under the IndiaAI Mission. Sarvam LLM demonstrates the feasibility of large-scale Hindifirst intelligence, achieving over 20% improvement on Indic benchmarks compared to global models.