ILA Annual Conference 2026 · Beyond Human Language: The Impact of AI on Linguistics John Jay College, City University of New York · New York City, USA · April 30 – May 2, 2026

Structural Silence

When AI Infrastructure Fails the World's Underrepresented Languages: A Case Study in Bengali, the Digital Divide, and the Hidden Cost of English-Centric Design

Avijit Roy John Jay College of Criminal Justice, CUNY
Proma Roy The City College of New York, CUNY

Bengali is the seventh most spoken language in the world — approximately 242 million native speakers and 285 million total speakers, a 1,400-year literary history, and UNESCO recognition of International Mother Language Day, commemorating those who died defending their right to speak it.

"285 million speakers.
Less than 0.5% of global web content.
A 67:1 training data gap."

When you try to build an AI tool that teaches in Bengali, you discover the entire infrastructure of modern machine learning was designed around a different language. That is not a coincidence. It is a policy outcome.

Bengali by the Numbers: Where Speakers and AI Resources Diverge

Speakers (Total incl. L2)
English
~380M
Bengali
~285M
Bengali total speakers (L1+L2): ~285M · ~4% of global population
Global Web Content
English
~49%
Bengali
<0.5%
Training Tokens (Billions)
English
~2,000B
Bengali
~30B
67:1 English-to-Bengali token ratio in major multilingual training corpora
Ethnologue 2025 · Pimienta, Research Outreach 2024 · Sangraha, Khan et al. 2024 · Common Corpus, Langlais et al. 2025 · BenLLM-Eval, Kabir et al. 2024

The Double Burden: Why Language of Instruction Matters

Native Language
🧠
Programming Concepts
- available capacity -
- available capacity -

Full working memory available for learning

Foreign Language (NO SUPPORT)
🧠
Programming Concepts
Language Processing
Language Processing

Cognitive Overload -> reduced retention -> dropout

Double Burden Linguistic processing competes with conceptual learning, creating a hard ceiling on technical mastery for speakers of unsupported languages.
Cognitive Load Theory (Sweller et al. 2011) · Roussel et al., Learning and Instruction, 2017 · Soosai Raj et al., ACM CHI, 2018

This is not only a technical problem.
It reflects deeper asymmetries in resources, infrastructure, and design.

🌐
<0.5%

The Web Presence Gap

Bengali speakers represent ~4% of the global population but account for less than 0.5% of global web content. AI training data is harvested from the web — so a language's online presence is not a cultural metric. It is a hard constraint on what AI systems can learn about it before training even begins.

Pimienta, Research Outreach 2024 · Ethnologue 2025 · ICLS Most Spoken Languages 2026
⚖️
67:1

The Training Token Deficit

The Sangraha corpus — one of the largest multilingual Indic datasets — allocates ~30 billion tokens to Bengali against ~2 trillion available for English in comparable corpora. Consistent, documented performance gaps across Bengali NLP benchmarks are the direct consequence. More data for English means better models for English. Always.

Khan et al. (Sangraha) 2024 · Langlais et al. (Common Corpus) 2025 · Kabir et al. (BenLLM-Eval) 2024 · Arxiv:2507.23248, 2025
🔡
+overhead

The Tokenization Penalty

Bengali's alphasyllabary script — where base characters are modified by diacritics and conjunct forms — is handled poorly by standard BPE and WordPiece tokenizers. Higher token fertility rates mean Bengali doesn't just need equal data. It structurally requires more data than English to achieve equivalent model performance.

Shahriar & Barbosa 2024 · Arxiv:2507.23248, 2025
📡
36.5%

The Connectivity Exclusion

Individual internet penetration in rural Bangladesh: 36.5% vs. urban 71.4% (BBS FY2024-25). This gap has widened year-over-year despite national investment. Cloud-dependent AI tools are, by design, tools for connected users. Offline-first architecture is not a technical compromise — it is the only design that actually reaches the learners who need these tools most.

Bangladesh Bureau of Statistics, ICT Access & Use Survey FY2024-25 · Daily Star, Jan 2, 2025
"Dataset scarcity is a structural barrier, not a researcher failure. Offline-first design is equity infrastructure, not a workaround. The linguistics community is uniquely positioned to name and contest the English-centric assumptions baked into AI pipelines — because those assumptions are, at their root, assumptions about whose language counts."

Reframe the Infrastructure Problem

  • Treat Bengali technical NLP benchmarks as primary research contributions — not preliminary work done before "the real research" begins.
  • Document dataset construction labor in publications. Normalize the cost of building foundational corpora for underrepresented languages.
  • Design multilingual evaluations that surface low-resource failure — not only reward high-resource performance. A model scoring 90% on English and 55% on Bengali is not a "multilingual" model.

Name the Assumptions Others Don't See

  • Recognize that "multilingual" AI models are not neutral. Their performance distributions reflect historical resource allocation decisions — not linguistic reality.
  • Advocate for offline-first AI design as a human-centered choice. Connectivity requirements are equity decisions. They should be named as such.
  • Linguistics has the critical vocabulary to analyze whose language is centered in AI systems — and whose is erased. That analysis is needed now more than ever.
  • Pimienta, D. (2024). Reliably exploring the presence of languages on the Internet. Research Outreach, (139), 58–61. https://doi.org/10.32907/RO-139-5856605838
  • Eberhard, D. M., Simons, G. F., & Fennig, C. D. (Eds.). (2025). Ethnologue: Languages of the world (28th ed.). SIL International. https://www.ethnologue.com/insights/ethnologue200/
  • Khan, M. S. U. R., Mehta, P., Sankar, A., Kumaravelan, U., Doddapaneni, S., B, S., G, V., Jain, S., Kunchukuttan, A., Kumar, P., Dabre, R., & Khapra, M. M. (2024). IndicLLMSuite: A blueprint for creating pre-training and fine-tuning datasets for Indian languages. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 9743–9752). Association for Computational Linguistics. https://aclanthology.org/2024.acl-long.843/
  • Langlais, P.-C., Rosas Hinostroza, C., Nee, M., Arnett, C., Chizhov, P., Jones, E. K., Girard, I., Mach, D., Stasenko, A., & Yamshchikov, I. P. (2025). Common Corpus: The largest collection of ethical data for LLM pre-training. arXiv. https://doi.org/10.48550/arXiv.2506.01732
  • Kabir, M., Islam, M. S., Laskar, M. T. R., Nayeem, M. T., Bari, M. S., & Hoque, E. (2024). BenLLM-Eval: A comprehensive evaluation into the potentials and pitfalls of Large Language Models on Bengali NLP. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 2238–2252). ELRA and ICCL. https://aclanthology.org/2024.lrec-main.201/
  • Bhowmik, S., Dipto, T. T., Islam, M. S., Hsu, S., & Reasat, T. (2025). Evaluating LLMs' multilingual capabilities for Bengali: Benchmark creation and performance analysis. arXiv. https://doi.org/10.48550/arXiv.2507.23248
  • Shahriar, A., & Barbosa, D. (2024). Improving Bengali and Hindi Large Language Models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 8719–8731). ELRA and ICCL. https://aclanthology.org/2024.lrec-main.764/
  • Sweller, J., Ayres, P., & Kalyuga, S. (2011). Cognitive load theory. Springer. https://doi.org/10.1007/978-1-4419-8126-4
  • Roussel, S., Joulia, D., Tricot, A., & Sweller, J. (2017). Learning subject content through a foreign language should not ignore human cognitive architecture: A cognitive load theory approach. Learning and Instruction, 52, 69–79. https://doi.org/10.1016/j.learninstruc.2017.04.007
  • Soosai Raj, A. G., Ketsuriyonk, K., Patel, J. M., & Halverson, R. (2018). Does native language play a role in Proceedings of the 23rd Annual ACM Conference on Innovation and Technology in Computer Science Education (ITiCSE '18) (pp. 21–26). Association for Computing Machinery. https://doi.org/10.1145/3197091.3197122
  • Bangladesh Bureau of Statistics. (2025). Quarterly report on ICT access and use survey 2024-25 (2nd Quarter). Government of the People's Republic of Bangladesh. View Official Report
  • Raihan, N., Jawad, M. A., Rahman, M. M., Ulfat, N., Gupta, P., Rahman, M. M., Karmaker, S., & Zampieri, M. (2025). Overview of BLP-2025 Task 2: Code generation in Bangla. In Proceedings of the Second Workshop on Bangla Language Processing (BLP-2025) (pp. 373–387). Association for Computational Linguistics. https://aclanthology.org/2025.banglalp-1.31/

If you use this work, please cite:

Paper Citation

            @article{roy2026structural_silence,
              title   = {Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages},
              author  = {Roy, Avijit and Roy, Proma},
              year    = {2026},
              journal = {SSRN Electronic Journal},
              publisher = {Elsevier BV},
              doi     = {10.2139/ssrn.6522858},
              url     = {https://doi.org/10.2139/ssrn.6522858}
            }
          

Poster Citation

          @misc{roy_2026_structural_silence_poster,
            author       = {Roy, Avijit and
                            Roy, Proma},
            title        = {Structural Silence: When AI Infrastructure Fails
                            the World's Underrepresented Languages —
                            Poster Presented at ILA 2026},
            month        = may,
            year         = 2026,
            publisher    = {International Linguistic Association (ILA)},
            doi          = {10.5281/zenodo.19991979},
            url          = {https://doi.org/10.5281/zenodo.19991979}
          }
          
Methodology Note
This ratio is derived by comparing the Bengali token allocation in Sangraha (~30B tokens; Khan et al. 2024) against the approximate English token count in Common Corpus (~2T tokens; Langlais et al. 2025). These are not the same corpus, and the comparison is intended to illustrate the order-of-magnitude disparity in available training data rather than an exact within-corpus measurement.