Which India Survives Translation? Narrative Homogenisation Across Indian Oral Traditions in LLMs
LLMs trained on English-dominated internet text tend to flatten diverse Indian storytelling traditions into a homogenized archetype, with cross-tradition similarity (0.52–0.66) far exceeding what genuine cultural distance would predict Prompting in regional languages (Hindi, Tamil, Bengali) consistently reduced fidelity to authentic traditions compared to English prompting, by up to 27 percentage points for Rajasthani and Bengali traditions The study examined three distinct Indian traditions—Raj
Analysis
TL;DR
- LLMs trained on English-dominated internet text tend to flatten diverse Indian storytelling traditions into a homogenized archetype, with cross-tradition similarity (0.52–0.66) far exceeding what genuine cultural distance would predict
- Prompting in regional languages (Hindi, Tamil, Bengali) consistently reduced fidelity to authentic traditions compared to English prompting, by up to 27 percentage points for Rajasthani and Bengali traditions
- The study examined three distinct Indian traditions—Rajasthani Pabuji epic, classical Tamil Sangam poetry, and Bengali folk tales—using Sentence-BERT embeddings and cosine similarity to measure reference drift and cross-tradition convergence
- This finding challenges assumptions that multilingual prompting inherently improves cultural authenticity, suggesting instead that it may elicit general cultural diversity rather than simulate narrow, lesser-documented traditions
- The pilot study offers a lightweight, scalable methodology complementary to large-scale human-annotation efforts for detecting cultural misrepresentation in LLM-generated content
Why It Matters
This research directly addresses a critical gap in AI safety and cultural representation: as LLMs are deployed globally, they risk erasing the nuance of non-Western oral and literary traditions by collapsing them into a single homogenized narrative. For practitioners building multilingual or culturally-aware systems, these findings suggest that simply switching to a regional language prompt does not guarantee authentic cultural output and may actively degrade fidelity to source traditions.
Technical Details
- Corpus: Authentic reference corpora collected for three Indian traditions—Rajasthani Pabuji epic (11 passages), classical Tamil Sangam poetry (21 passages), and Bengali folk tales (10 passages)
- Models tested: Claude Sonnet and Gemini, prompted with 54 generation requests across three prompt types per tradition (generic, culturally specific, and regional-language)
- Methodology: Sentence-BERT embeddings combined with cosine similarity to compute two metrics: reference drift (output proximity to its own tradition's authentic texts vs. others) and cross-tradition convergence (similarity of outputs across traditions)
- Key quantitative finding: Cross-tradition similarity ranged from 0.52 to 0.66, substantially higher than the distance between authentic traditions would predict; regional-language prompting reduced fidelity by up to 27 percentage points compared to English prompting for Rajasthani and Bengali traditions
- Positioning: Framed as a lightweight, scalable complement to large-scale human-annotation studies, part of a broader doctoral research program on Indian cultural misrepresentation in LLMs
Industry Insight
- Organizations deploying LLMs for culturally-specific content generation should not assume multilingual prompting automatically improves authenticity; rigorous evaluation against ground-truth cultural corpora is essential before relying on regional-language outputs
- The counterintuitive finding that regional-language prompting degrades fidelity suggests current LLMs may conflate broad cultural tropes with tradition-specific nuance when prompted in non-English languages—a risk worth monitoring as multilingual capabilities expand
- This pilot methodology (embedding-based drift and convergence measurement) offers a scalable template for auditing cultural representation across other underrepresented traditions, enabling proactive rather than reactive cultural safety assessments
Disclaimer: The above content is generated by AI and is for reference only.