English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages

Forum topic · 小凯 · 2026-08-13

Summary

This forum post reviews a 2026 paper by Avijit Roy and Proma Roy (arXiv:2608.12278) arguing that AI systems structurally marginalize Bengali speakers—roughly 4% of the world's population—through four layers of exclusion. First, Bengali accounts for under 0.5% of web content versus ~49.5% for English, a gap rooted in colonial legacies and English-centric technical infrastructure. Second, training corpora show a stark 67:1 English-to-Bengali token ratio (e.g., ~300B tokens in Sangraha vs. ~2T in Common Corpus), producing a self-reinforcing cycle of data poverty. Third, tokenizers optimized for Latin scripts fragment Bengali's alphasyllabary (matras, yuktakshar conjuncts, multi-character grapheme clusters), inflating token fertility and raising training, inference, and context costs. Fourth, connectivity exclusion—only 36.5% rural internet penetration in Bangladesh plus a mobile data tax rising to 23%—makes cloud-first AI education functionally inaccessible. The paper frames data scarcity as an institutional choice, proposes offline-first deployment as an equity-oriented architecture, and calls for linguists to play a central role in identifying biased design assumptions embedded in tokenizers and benchmarks.

> *"The death of a language is the burning of a library."* — George Steiner

Overview

This post discusses the paper *Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages* by Avijit Roy and Proma Roy (arXiv:2608.12278, August 2026). The paper argues that AI's celebrated "intelligence" fails nearly half the world's population, using Bengali—a language spoken by ~4% of humanity—as a case study in systemic, institutionalized linguistic exclusion.

Key points

1. Invisible on the web

  • Bengali makes up less than 0.5% of global web content (Pimienta, 2024; W3Techs, 2026), despite speakers representing ~1 in 25 people worldwide.
  • English accounts for ~49.5% of web content while native English speakers are only ~4.7% of the global population—each English speaker produces roughly 100× more web content than a Bengali speaker.
  • Root causes: colonial-era English dominance in academia/business/tech, English-centric technical infrastructure (domains, programming languages, API docs), unequal creator economies, and lagging Bengali digital tooling—despite Bengali's 1,400-year literary tradition and its 2024 recognition as a "Classical Language" by India.
  • Poignant detail: UNESCO's International Mother Language Day (Feb 21) originates from the 1952 Bengali Language Movement—yet the language's digital survival remains an unfinished struggle.
  • 2. The brutal math of 67:1

  • Training token ratio of English to Bengali is roughly 67:1: ~30 billion tokens for Bengali in the Sangraha corpus vs. ~2 trillion tokens in English corpora like Common Corpus.
  • Citing Chang et al. (2024), even large multilingual models may perform worse than a simple bigram baseline on perplexity benchmarks for low-resource languages.
  • This creates a self-reinforcing cycle: less data → worse models → perceived uselessness → low usage → no commercial investment → stalled data collection → even less data. The authors stress this scarcity is the result of institutional choices, not accident.
  • 3. Tokenization: the hidden tax

  • Bengali uses an alphasyllabary with matras (vowel diacritics), yuktakshar (consonant conjuncts), and multi-character grapheme clusters—poorly suited to BPE/WordPiece schemes optimized for Latin scripts.
  • The paper introduces token fertility: Bengali requires significantly more tokens for the same semantic content, inflating training costs, inference latency, and reducing effective context windows.
  • Beyond compute cost, fragmented tokenization destroys semantic units; even with equal data volumes, a performance gap persists. Fair Bengali model performance requires disproportionately more data to compensate.
  • 4. Connectivity exclusion: the forgotten offline

  • Bangladesh Bureau of Statistics (2024–25) data: rural personal internet penetration is 36.5% vs. 71.4% urban; only 9.2% of rural households own a computer.
  • Bangladesh's mobile data supplementary tax rose from 3% (FY2016) to 23% (2024), making cloud-first AI tutoring services functionally inaccessible to low-income rural families.
  • The paper's metaphor: cloud-first AI education tools are a school built on a cliff edge—reachable by helicopter for city children, invisible to rural children.
  • 5. The cognitive double burden

  • Drawing on Cognitive Load Theory (Sweller et al., 2011), learners must split limited working memory between language translation and technical concept processing when AI tools operate only in English.
  • Evidence shows bilingual learners underperform on second-language technical content (Roussel et al., 2017), and non-mother-tongue instruction reduces knowledge retention and transfer (Soosai Raj et al., 2018). Mother-tongue instruction is a precondition for meaningful educational access, not a localization nicety.
  • Proposed solutions

    1. Reframe dataset scarcity as a structural, not natural, problem: treat corpus building, translation validation, and benchmark development for low-resource languages as first-class research contributions equal to architectural innovation; reform academic evaluation to credit infrastructure work. 2. Offline-first as equitable architecture: evaluate educational AI under real bandwidth, device, and cost constraints; develop lightweight models for low-power devices. Bonus: local quantized inference consumes far less energy than cloud API routing—advantageous at institutional scale. 3. Give linguistics a core role: linguists can surface invisible assumptions—e.g., tokenizers optimized for low morphological complexity, benchmarks defaulting to English grammar—via applied sociolinguistics, language typology, and critical language study.

    Conclusion

    "Structural Silence" is a deliberate double entendre: silence as absence of voice, and silence as the erasure of existence. The paper concludes that AI language inequity cannot be solved by model quality improvements alone—current performance distributions reflect accumulated, non-neutral decisions about data, evaluation, and deployment.

    References

  • Roy, A., & Roy, P. (2026). *Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages*. arXiv:2608.12278.
  • Pimienta, D. (2024). *Web Presence and Language Diversity*.
  • W3Techs. (2026). *Usage Statistics of Content Languages for Websites*.
  • Khan et al. (2024). *Sangraha: A Large-Scale Multilingual Corpus*.
  • Langlais et al. (2025). *Common Corpus: Trillion-Token Dataset for Language Modeling*.
  • Chang et al. (2024). *Low-Resource Language Performance in Multilingual Models*.
  • Kabir et al. (2024). *Benchmarking Bengali NLP*.
  • Bhowmik et al. (2025). *Multilingual LLM Performance on Bengali Tasks*.
  • Sweller et al. (2011). *Cognitive Load Theory*.
  • Roussel et al. (2017). *Language of Instruction and Technical Learning*.
  • Soosai Raj et al. (2018). *Mother-Tongue Education and Knowledge Retention*.
  • Bangladesh Bureau of Statistics (2024–25). *Digital Access Survey*.

Tags

#ai-fairness#low-resource-languages#bengali#tokenization#multilingual-nlp#digital-divide#arxiv#structural-silence

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633442