This forum post introduces and analyzes the paper *Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages* by Avijit Roy and Proma Roy (arXiv:2608.12278), focusing on how AI systems systematically exclude Bengali speakers—nearly 4% of the global population.
Key points
1. Web content invisibility
- Bengali represents under 0.5% of global web content (Pimienta, 2024; W3Techs, 2026), despite ~4% of humanity speaking it—roughly a two-orders-of-magnitude gap.
- English holds ~49.5% of web content with only ~4.7% native speakers, meaning each English speaker produces ~100x more web content than a Bengali speaker.
- Root causes: colonial-era linguistic hierarchy, English-centric technical infrastructure, unequal creator economies, and lagging Bengali digitization tools. The post notes the irony that International Mother Language Day (Feb 21) originates from the 1952 Bengali Language Movement.
- English-to-Bengali training token ratio is approximately 67:1: Sangraha provides ~30 billion Bengali tokens vs. ~2 trillion English tokens in Common Corpus.
- Chang et al. (2024) found large multilingual models can underperform simple bigram baselines on low-resource language perplexity benchmarks.
- The post highlights a self-reinforcing cycle: scarce data → poor models → low usage → low investment → even scarcer data—driven by institutional choices in funding, publishing, and market priorities, not accident.
- Bengali's alphasyllabary (matras, yuktakshar conjuncts, multi-character grapheme clusters) clashes with Latin-optimized BPE/WordPiece schemes.
- Higher token fertility means Bengali text consumes more tokens for identical semantic content, inflating training/inference costs, shortening effective context windows, and degrading attention quality—so parity requires disproportionately more data.
- Bangladesh Bureau of Statistics (2024–25): rural individual internet penetration 36.5% vs. 71.4% urban; only 9.2% of rural households own a computer.
- Mobile data supplementary tax rose from 3% (FY2016) to 23% (2024), making cloud-first AI tools functionally inaccessible offline.
- Citing Cognitive Load Theory (Sweller et al., 2011), Roussel et al. (2017), and Soosai Raj et al. (2018): non-mother-tongue instruction measurably reduces comprehension, retention, and transfer—mother-tongue AI education is a prerequisite, not a localization feature.
- Roy, A., & Roy, P. (2026). *Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages*. arXiv:2608.12278.