Research Summary: Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages

Research identifies systemic failures in Artificial Intelligence (AI) infrastructure that significantly disadvantage speakers of underrepresented languages. These failures, including training data limitations, tokenization issues, and evaluation benchmark biases, exist before model deployment, creating 'structural silence.' The case study of Bengali, a widely spoken language with limited web presence, highlights these challenges, particularly in AI-assisted education in low-connectivity settings. This indicates a critical gap between the promise of AI for educational access and its practical limitations for a substantial global population.

Why it matters

This research underscores a fundamental challenge to the equitable deployment and beneficial impact of AI technologies globally. It highlights that current AI development paradigms may exacerbate existing societal inequalities rather than mitigate them, particularly affecting linguistic diversity and access to essential services like education.

Key insights

• AI infrastructure systematically disadvantages speakers of underrepresented languages prior to model training.
• Critical components like training corpora, tokenization schemes, and evaluation benchmarks contribute to this disparity.
• The issue is termed 'structural silence,' signifying a built-in failure within AI systems for certain linguistic groups.
• Bengali, despite being a widely spoken language, represents less than 0.5% of global web content, contributing to a severe 'web presence gap'.
• These failures are particularly impactful for AI-assisted education in environments with low connectivity.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.12278