Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages

Synthesis: Roy & Roy (2026) argue that the infrastructure of AI — training corpora, tokenization, benchmarks, deployment architectures — systematically disadvantages speakers of underrepresented languages before a model is trained, reframing dataset scarcity as a structural barrier rather than an isolated technical limitation. Using Bengali as a case in AI-assisted education, they document four interlocking failures: a web-presence gap (<0.5% of global content for ~4% of the population), a 67:1 English↔Bengali training-token deficit, a tokenization penalty from the alphasyllabary script, and connectivity exclusion (36.5% rural vs 71.4% urban internet penetration). They position offline-first design as an equity-oriented infrastructure strategy. The work connects to Equity, Language Learning, and Digital Divide debates in educational AI.

Key Findings

1. Underrepresented languages are excluded from AI infrastructure through structural, design-level decisions — not explicit policy — and this exclusion compounds across four layers: web presence, training tokens, tokenization, and deployment connectivity.
2. Bengali represents roughly 4% of the global population but less than 0.5% of global web content, while English holds ~49.5% of web content — a web-presence gap that precedes any model training.
3. Major multilingual corpora allocate tokens at a 67:1 English-to-Bengali deficit, and Bengali's alphasyllabary script compounds the data shortfall by forcing higher token fertility under standard Latin-script tokenizers, so equal data volumes still leave a performance gap.
4. Rural learners face a connectivity exclusion (36.5% individual internet penetration vs 71.4% urban) that makes cloud-dependent AI tools functionally inaccessible, turning native-language, offline-first design into a prerequisite for educational access rather than a convenience.

Four Interlocking Infrastructure Failures

The paper identifies four structural barriers that compound to exclude underrepresented languages from AI-assisted education:

• Web presence gap: Bengali accounts for under 0.5% of global web content despite representing nearly 4% of the global population, while English commands approximately 49.5%. Because Large Language Models (LLMs) training corpora are largely harvested from web crawls like Common Crawl, a language's digital presence constrains its representation in training data.
• Training-token deficit: a 67:1 deficit between English and Bengali in major multilingual corpora. The Sangraha corpus allocates ~30B Bengali tokens against ~2T English tokens in Common Corpus, and inclusion in a multilingual set does not guarantee competence — large models like XGLM and BLOOM can underperform bigram baselines on low-resource perplexity benchmarks.
• Tokenization penalty: Bengali's alphasyllabary script compounds the data deficit through higher token fertility. Standard large language model tokenizers (BPE, WordPiece) were optimized for Latin scripts, and matras and conjuncts (yuktakshar) fragment into noisy subword clusters that disrupt linguistic units.
• Connectivity exclusion: individual internet penetration is 36.5% in rural areas versus 71.4% in urban areas; household access is not individual access, and rising data costs (mobile data duty up from 3% to 23%) price out lower-income rural learners.

These failures reflect longstanding resource-allocation decisions, institutional priorities, and design defaults that did not center underrepresented languages in mainstream AI development. The paper frames this cumulative exclusion as structural silence.

Reframing Scarcity as Structure

The authors argue dataset scarcity should be understood as a structural barrier rather than an isolated technical limitation. Treating scarcity purely as a data-volume problem obscures its structural origins in funding patterns, publication incentives, and research priorities that directed resources toward high-resource languages. When researchers build Bengali AI tools, they perform foundational infrastructure work — curating datasets, validating translations, building evaluation benchmarks — that the field has historically declined to fund as primary research. The absence of a Bengali programming-education corpus is not natural scarcity but the cumulative effect of institutional decisions about which outputs deserve recognition. This is a core concern of Equity and the Digital Divide literature.

Educational Consequences: The Double Burden

For Bengali-speaking learners using AI-assisted programming tools, the four failures converge in a compounded cognitive burden. Learning to program, and especially debugging, is cognitively demanding regardless of first language; when an AI tool delivers explanations in English — as any general-purpose model will, given the training-data distribution — learners with limited English proficiency must process linguistic content and technical content concurrently, frequently exceeding working-memory capacity under cognitive-load theory. Empirical work (e.g. Roussel et al., with 294 higher-education students) shows foreign-language content yields lower outcomes than native-language instruction, and bilingual programming instruction outperforms English-only instruction. Native-language AI explanation is thus a prerequisite for meaningful access in multilingual contexts, not an enhancement — a central tension for AI in Education design in low-resource, global-south settings.

Offline-First as Equity Architecture

The authors recommend treating offline-first design as an equity-oriented infrastructure strategy for AI-assisted education in low-connectivity environments rather than a secondary technical compromise. Local inference on quantized, parameter-efficient models (e.g. QLoRA, LoRA) has made on-device deployment feasible, extending reach to users excluded by cloud connectivity assumptions. This choice also carries a Sustainability dividend — local inference on a quantized 7B-parameter model consumes a fraction of the energy of routing every query through a remote API. Evaluation frameworks that test only high-connectivity performance implicitly validate tools that are inaccessible to large portions of their stated target populations. This reframing connects to Accessibility, Inclusive Learning, and Bias Mitigation: design choices that assume connectivity are themselves a form of exclusion.

What this means for practice

• Policymakers. Fund low-resource language infrastructure — datasets, benchmarks, and evaluation protocols — as primary research rather than preliminary support work, since Bengali holds under 0.5% of global web content against a 67:1 English-to-Bengali training-token deficit.
• Institutions. Require offline-first capability from any AI tool intended for rural learners, where individual internet penetration is 36.5% versus 71.4% urban, and evaluate it under the bandwidth, device, and cost constraints of the communities it serves.
• Institutions. Treat explanations in the learner's own language as a prerequisite for access rather than an enhancement, given evidence that foreign-language instruction yields lower outcomes for higher-education students.
• Researchers. Put linguistic analysis at the center of AI development and evaluation so that tokenization assumptions — such as why standard BPE fits Bengali's modification structure poorly — are named rather than hidden by English serving as the reference standard.

Limitations

• The paper is a theoretical analysis built on a single-language case, and the authors describe its contribution as explicitly "not a new benchmark or deployment system," so it cannot support claims about learning outcomes for Bengali-speaking students.
• Its quantitative claims come from secondary sources rather than collected data: the web-presence gap (under 0.5% of content for roughly 4% of the global population), the 67:1 token deficit, and the connectivity figures (36.5% rural vs 71.4% urban) rest on corpus statistics and national survey reports that date.
• Bengali is a favorable case because of its demographic scale, literary history, and active NLP research community; the authors state that smaller and less-resourced languages face considerably more severe structural challenges, so the four-failure account is a framework to test, not a cross-linguistic finding.

Citation

Roy, A., & Roy, P. (2026). Structural silence: When AI infrastructure fails speakers of underrepresented languages.