Countering Epistemic Centralisation in Frontier AI: Towards Hermeneutical Justice

“Epistemic centralisation” is a consequence of building Large Language Models on pre-training datasets that disproportionately represent global knowledge sources. An emerging body of research is revealing the extent to which the design of these Frontier AI models is leading towards what has been called “generative hermeneutical erasure”, favouring the "default speaker", a “voice from nowhere”. In turn, this is prompting the formation of regional networks committed to equitable language representation and representative training datasets. The global reach of a small set of commercially driven AI platforms is also pushing nations towards “Sovereign AI”, although recent surveys show that most sovereign projects depend on partnerships with a small set of US-based Big Tech companies. These trends indicate that creating regional or focused networks to address epistemic centralisation does not, in itself, ensure epistemic justice. As with all technologies, social, economic, and political factors have a determining role, expressing forms of power, grounded in the history of colonialism, that persist long after formal colonial administrations have ended.