Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages

Forum topic · 小凯 · 2026-08-13

Summary

This article discusses a 2026 paper by Avijit Roy and Proma Roy examining how AI systems systematically marginalize Bengali, spoken by roughly 4% of the world's population. Four overlapping forms of 'structural silence' are identified: underrepresentation on the web (Bengali constitutes under 0.5% of online content despite its speaker base), severe training-data imbalance (an estimated 67:1 English-to-Bengali token ratio in major corpora), tokenization inefficiency caused by alphasyllabary script features that inflate token fertility, and connectivity-based exclusion (only 36.5% of rural Bangladeshis have personal internet access, with mobile data surcharges rising to 23%). The authors argue that low-resource language poverty reflects institutional choices in research funding, publishing incentives, and commercial priorities rather than natural scarcity. They propose reframing data scarcity as a structural problem, treating offline-first deployment as an equity-oriented design principle, and embedding applied linguistics into AI development to surface hidden assumptions.

Overview

This article examines "Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages," a 2026 paper by Avijit Roy and Proma Roy (arXiv:2608.12278). The authors argue that AI systems marginalize Bengali through four overlapping mechanisms, each producing compounding disadvantages for roughly 250 million speakers.

Key Points

1. Web Presence Gap

Bengali accounts for under 0.5% of global web content (Pimienta, 2024; W3Techs, 2026), despite being spoken by ~4% of the world's population.
English, by contrast, makes up ~49.5% of web content while its native speakers also constitute ~4.7% of the global population.
Per-capita English content production is roughly 100× that of Bengali, rooted in colonial legacy, English-centric technical infrastructure, and unequal content-economy incentives.
International Mother Language Day (February 21) commemorates the 1952 Bengali Language Movement; the paper frames ongoing digital underrepresentation as an unfinished struggle.

2. Training-Data Imbalance

Major multilingual corpora exhibit an estimated 67:1 English-to-Bengali token ratio.
Sangraha allocates ~30 billion tokens to Bengali; Common Corpus and similar English datasets reach ~2 trillion tokens.
Chang et al. (2024) show that large multilingual models can underperform a simple bigram baseline on perplexity benchmarks in low-resource languages, indicating that model complexity cannot compensate when data poverty crosses a critical threshold.
A self-reinforcing cycle is identified: poor model performance → low user adoption → minimal commercial investment → stalled data collection.

3. Tokenization Inefficiency

Bengali uses an alphasyllabary featuring matras, conjuncts (yuktakshar), and multi-character grapheme clusters.
Standard BPE/WordPiece tokenizers, optimized for Latin orthographies, fragment Bengali text into many sub-semantic pieces, inflating token fertility.
Higher token counts raise training cost, slow inference, shrink effective context windows, and dilute attention—making the script a structural handicap built into tokenizer design.

4. Connectivity-Based Exclusion

Bangladesh Bureau of Statistics (2024-25) reports individual internet penetration of 36.5% in rural areas vs. 71.4% in urban areas; only ~9% of rural households own a computer.
Cloud-first AI education tools assume stable WiFi, generous data plans, and cloud-service affordability—assumptions that fail for most rural users.
Mobile data surcharges rose from 3% (FY2016) to 23% (2024), effectively making cloud AI functionally inaccessible.
The authors reframe offline-first deployment not as a degraded fallback but as an equity-oriented infrastructure strategy: evaluation under bandwidth and device constraints, lightweight on-device models, and offline-first interaction design also yield sustainability benefits through lower energy consumption.

5. Cognitive Load Consequences

Drawing on Cognitive Load Theory (Sweller et al., 2011), the paper shows that bilingual learners processing technical content in a second language experience measurable performance and retention deficits (Roussel et al., 2017; Soosai Raj et al., 2018).
Mother-tongue instruction is framed as a prerequisite for meaningful educational access, not a localization enhancement.

Proposed Responses

1. Reframe data scarcity as a structural—not natural—condition: elevate corpus construction, validation, and benchmark development to first-class research contributions; reform academic evaluation; sustain institutional funding for underrepresented languages. 2. Offline-first as fairness infrastructure: evaluate AI education tools under real connectivity and device constraints; build small, efficient on-device models. 3. Center applied linguistics in AI development: use sociolinguistics to identify whose grammars standard tools parse, typology to compare structural differences, and critical language studies to surface whose needs enter system design.

Conclusion

The paper's "Structural Silence" framing deliberately echoes the political weight of silence in Bengali linguistic history. Performance disparities across languages reflect cumulative institutional decisions about funding, evaluation, and deployment—decisions that are not neutral. Closing the gap requires treating language equity as infrastructure, not as a post-hoc localization task.

Key References

Roy, A., & Roy, P. (2026). *Structural Silence*. arXiv:2608.12278
Pimienta, D. (2024). *Web Presence and Language Diversity*
W3Techs (2026). *Usage Statistics of Content Languages*
Khan et al. (2024). *Sangraha Corpus*
Langlais et al. (2025). *Common Corpus*
Chang et al. (2024). *Low-Resource Language Performance in Multilingual Models*
Sweller et al. (2011). *Cognitive Load Theory*
Roussel et al. (2017); Soosai Raj et al. (2018)
Bangladesh Bureau of Statistics (2024-25). *Digital Access Survey*