Building a Multilingual Digital Language Public Good Stack

The project identifies a genuine gap in multilingual AI safety infrastructure and presents a thoughtful vision for an open, knowledge-graph-grounded evaluation ecosystem for under-resourced languages. Framing multilingual safety infrastructure as a digital public good is a compelling perspective with long-term value.

The project is addressing a genuine concern of multilingual safety gap as an AI infrastructure problem. Two very different languages Kannada/Khmer are chosen to build a knowledge graph and a KG grounded evaluation pipeline, demonstrated on a "vertical slice" in the electoral and civic domain. However, the implementation is missing, so the project remains inconclusive in the real world.

This blueprint makes a strong conceptual contribution by framing multilingual safety gaps for under‑served Asian languages as an infrastructure problem and proposing a “Digital Language Public Good Stack” that combines open corpora, a Wikidata‑aligned knowledge graph, and a knowledge‑graph‑grounded safety evaluation pipeline for Kannada and Khmer, focused on electoral/civic information with a planned extension to socio‑cultural harms.

The architecture, threat model, and stakeholder use‑cases are clearly articulated and draw thoughtfully on African benchmarks, multilingual safety work, and digital public goods frameworks, making this highly relevant for Global South AI safety, but empirical execution is explicitly deferred—this is a design document without implemented experiments, metrics, or results, so methodological rigour cannot yet be assessed beyond planning quality.

The writing is well structured and accessible, with clear limitations and future‑work steps, and the piece would be even stronger if it included a small “toy” implementation slice (e.g., a mini KG plus a handful of evaluation prompts and concrete failure examples) to demonstrate feasibility and provide early evidence that the proposed stack yields actionable safety signals.