Sovereign AI vs AI Polyglots Lecture

​Lecture about paper 'Sovereign AI vs AI Polyglots'

​Temporary link: https://www.researchgate.net/publication/396680116_Sovereign_AI_vs_AI_Polyglots

The global AI landscape is dominated by a United States-China duopoly in model development, data, and compute resources, raising concerns about both cultural and infrastructural hegemony. This paper contrasts two paradigms: AI polyglots, which are large language models (LLMs) optimized for broad multilingual fluency but limited cultural depth and primarily developed by the two major powers, and sovereign AI, which aligns models with a nation's language, culture, and computing infrastructure. Drawing on insights from cognitive psychology and analogies with human perceptual and inferential biases, we suggest that epistemic anchoring in training data shapes model priors. This perspective underscores the importance of sovereign AI for smaller nations, enabling systems that are attuned to local contexts rather than inheriting American or Chinese worldviews and decision-making patterns. We propose a dual-axis framework for assessing AI sovereignty: epistemic sovereignty, achieved through culturally grounded training data that fosters intentional, beneficial biases aligned with local values, norms, and knowledge traditions; and operational sovereignty, achieved through control over localized compute and inference infrastructure. To make these dimensions measurable, we introduce metrics that quantify parameter allocation per language and assess the share of culturally rooted data in the corpus. We further formalize cultural misalignment as a conditional distribution shift, drawing on classical machine learning theory. Finally, we emphasize that sovereign AI need not be monolingual. In multilingual nations and supranational initiatives, achieving genuine cultural understanding requires models that represent all linguistic communities and capture interactions between dominant and minority languages.