Governance & AI

The Digital Linguistic Bias (DLB) and AI Governance

What happens when the AI systems shaping access to knowledge don't speak the language of over 650 million people with sufficient fidelity? This talk starts from an uncomfortable premise: Large Language Models (LLMs) are not linguistically neutral. Most are built on an English-language base that can account for up to 90% of their document corpus, with Spanish content derived largely from automated translation—rendering entire dialectal varieties invisible and distorting the representation of real speaker communities.
Mariana Hungría analyzes the concept of Digital Linguistic Bias (DLB), introduced by Muñoz-Basols, Palomares Marín, and Moreno Fernández (2024), as a conceptual framework for understanding the linguistic hybridity that AI generates at both interlinguistic level—the structural weight of English over other languages—and intralinguistic level—the hierarchization among varieties of Spanish itself. This bias is not a minor technical flaw: it is a governance issue.
The talk connects the findings of the reference paper to broader implications for those working in AI safety and governance: what it means that the world's most advanced systems fail to adequately represent the global majority of their users, what institutional mechanisms are—or are not—being activated across the Spanish-speaking world, and how the absence of a coordinated strategy between countries reproduces and amplifies existing epistemic inequalities.
A session designed for those who understand that governing AI also means governing the language through which that AI thinks.

Mariana Hungría
Linguist and Independent Researcher in AI Safety and Governance
Researcher specializing in the intersection of language, AI safety, and governance. Her work is built on a conviction the field often overlooks: that access to technical research shouldn't depend on the language in which you think. To that end, she co-leads translation projects that bring the most relevant AI debates to Spanish-speaking communities, and co-organizes the BlueDot Impact AI Governance Club, a training and networking space for those who want to contribute actively to more informed, more equitable AI governance.