A New Research Blueprint Asks: Can AI Actually Understand India's Languages, or Just Process Them?

A newly published paper, "Rethinking Indic AI from a Lens of Cultural Heritage Preservation," argues that the dominant approach to building AI for India's languages is missing something fundamental: culture itself. Posted to arXiv, the paper reframes Indic AI not as a scaling problem to be solved with more data and bigger models, but as a preservation problem, one where the stakes include the survival of scripts, dialects, and knowledge systems that predate the computing era by millennia.

The authors ground their argument in the sheer structural difficulty of Indian languages. Rich morphology, complex scripts, intricate grammar rules, diglossia (the coexistence of formal and colloquial registers), and heavy dialectal variation across regions all combine to make India's linguistic landscape uniquely resistant to the tokenization and training assumptions baked into most large language models. These aren't edge cases to be patched later, the paper suggests, but core design constraints that Indic foundation models need to be built around from the start.

From resource gaps to "Culture Sensing"

Much of the paper traces the history of Indic Natural Language Processing as a field, walking through key milestones, shifts in methodology, and the resource-creation efforts that have tried to close the representation gap for low-resource Indian languages. That historical survey sets up the paper's central proposal: a new research direction the authors call "Culture Sensing."

Rather than treating cultural context as metadata to be tacked onto a model's outputs, Culture Sensing is framed as a hermeneutic approach, one rooted in interpretation and meaning-making rather than pure statistical pattern matching. The goal is twofold: equitable performance across languages that today receive far less training data than English or Mandarin, and outputs that are genuinely culturally coherent rather than technically fluent but contextually hollow.

BharatiyaGPT as a working example

The paper points to BharatiyaGPT as a concrete instance of this philosophy in practice. According to the research, the system is being applied to digitize fragile physical artifacts and to assist in deciphering endangered languages, tasks that sit closer to archival science and linguistics than to conventional chatbot development. Framed this way, BharatiyaGPT becomes less a product pitch and more a case study in what happens when AI development is explicitly tied to cultural revitalization rather than commercial deployment alone.

Why this matters beyond India

The broader implication is that language coverage metrics, the kind AI labs often cite when claiming support for "100+ languages," may be a poor proxy for whether a model actually respects the linguistic and cultural realities of its speakers. India's scale (thousands of dialects, dozens of scripts, and a documented history stretching back further than most written languages) makes it a demanding test case. If Culture Sensing or similar frameworks prove workable there, they could offer a template for other regions where AI development has so far treated cultural nuance as an afterthought rather than a design requirement.