The paper titled, ‘Rethinking Indic AI from a Lens of Cultural Heritage Preservation’, authored by Aparna Madva, Sharath Srivatsa, Srinath Srinivasa and Tulika Saha, argue that while AI promises wider access to information, it also poses a serious threat to India’s linguistic and cultural plurality by reinforcing dominant narratives and under-representing regional languages and their distinct world-views.

The study identifies Tamil as a compelling example of the limitations of present-day large language models (LLMs). It argues that the language embodies far more than grammar or vocabulary, carrying centuries of literary traditions, philosophical thought, social practices and cultural interpretations that cannot be faithfully reproduced through translation-driven AI systems.

According to the paper, the problem stems from the way most foundation models are developed. Existing LLMs are largely trained on English-language resources and translated datasets that disproportionately reflect urban contexts and dominant linguistic communities. As a result, they frequently overlook the linguistic diversity and cultural nuances embedded in languages such as Tamil, producing responses that may be syntactically accurate but culturally incomplete or misleading.

Researchers argue that Tamil presents challenges that mainstream AI systems are yet to overcome. Its rich morphology, agglutinative word formation, phonetic writing system, flexible sentence structure and coexistence of literary and spoken forms demand language models capable of understanding context rather than merely predicting words. They caution that current AI architectures remain ill-equipped to preserve these linguistic characteristics, increasing the risk of cultural homogenisation as AI-generated content becomes more widespread.

The paper further warns that dialects, colloquial speech and orally transmitted knowledge, particularly those preserved in rural and indigenous communities, remain inadequately represented in AI training data. Such gaps, the authors say, could gradually marginalise regional identities as automated systems increasingly shape education, governance, digital communication and knowledge dissemination.

Tracing the evolution of Indic Natural Language Processing over several decades, the study acknowledges advances in language technologies, including IndicBERT, MuRIL, IndicTrans2 and other foundation models. However, it adds that improvements in computational performance have not translated into meaningful representation of India’s cultural diversity. The authors argued that technological progress had outpaced efforts to preserve the distinct world-views embedded in regional languages, leaving a critical gap in the development of inclusive AI.

To address this imbalance, researchers propose a new framework called ‘Culture Sensing’, which seeks to embed indigenous knowledge systems, regional dialects and culturally grounded reasoning into future AI models. Rather than treating Tamil and other Indian languages as datasets for translation alone, the framework advocates building AI capable of recognising the cultural contexts in which these languages evolved.

Without such a shift, the paper argues, AI could emerge not merely as a technological tool but as a force that steadily weakens the linguistic and cultural heritage it is expected to serve.