Answers are grounded in this paper and cite their sources.
The claim that large language models are multilingual describes their user interface. Their internal architecture remains monolingual. Commercial systems generate text across dozens of languages, encouraging institutions to treat them as neutral global infrastructure. Beneath that polyglot surface, modern architectures operate as an English-centric monopoly. Training corpora drawn from the commercial web concentrate resources in English, rendering the planetary majority of languages computational noise. Inside the network, intermediate processing layers translate non-English queries into English representations before projecting answers back into the user's language. This architectural monopoly produces compounding inequalities. Non-Western scripts are shattered into fragmented processing units, imposing economic and memory penalties on non-English users. Safety guardrails trained in English collapse when prompts are translated into lower-resource tongues. Meanwhile, translated benchmarks certify capability by testing foreign fluency in Western cultural norms. Sovereign AI initiatives in India, Aotearoa New Zealand, and Africa reject this extractive pipeline. They demonstrate that linguistic self-determination requires native pretraining and enforceable regulatory parity. Keywords: Linguistic Imperialism, Multilingual Large Language Models, Latent Language Hypothesis, Internal Translation, Tokenization Disparity, Cross-Lingual Safety, Benchmark Parity, Indigenous Data Sovereignty, CARE Principles, Sovereign AI Infrastructure, Bhashini, BharatGen, Te Hiku Media, Masakhane, Algorithmic Monoculture, Computational Equity, Regulatory Certification, Model Auditing, Data Extraction, Language Reclamation
No takes yet. Share an insight, caveat, or question.
Jefferson et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: