On 6 March 2025, in the culture and language session of the “Rise of Asia in Global History and Perspective” conference at the University of Le Havre, France, Mr. Zandberg presented the study “Evaluating Digital Translation Services for Low- and Medium-Resource Languages in Public Administration.” The presentation discussed the pitfalls and promise of AI-powered translation for marginalized languages, connecting the topic to general issues of digital inclusion, multilingual governance, and the use of technology for cultural preservation.
As the Executive Director of the Global Diplomatic Council and a Professor of International Relations at the Geneva Nations Institute, Mr. Zandberg’s research interest lies at the crossroads of language policy, digital governance, and AI-driven translation technology. Bilingual public administration in the German regions where Frisian is a low-resource minority language with weak digital infrastructure for machine translation, was one of the key case studies in the presentation.
The Digital Language Divide: A Challenge for Public Governance
The presentation began by pinpointing a critical issue: while AI translation has revolutionized communication for high-resource languages (i.e., English, Mandarin, and Spanish), it has left hundreds of low-resource and minority languages—spoken by tens of millions—with inadequate digital representation. This deficit has a direct bearing on governance, accessibility to public services, and the ability of linguistic communities to contribute meaningfully to digital society.
On the basis of the Global Digital Compact’s goal to bridge digital divides, to become more inclusive, and to address AI governance, the presentation emphasized that there is a role that governments and policymakers can play so that AI translation technology spreads its benefits to all language groups and not just those with the largest data sets.
Current Status of AI Translation in Low-Resource Languages
The presentation gave an overview of major AI-powered translation services, such as:
- Google Translate & Microsoft Translator, which in recent years have introduced dozens of new languages but continue to battle with low-resource language precision.
- Meta’s No Language Left Behind (NLLB-200), a cutting-edge multilingual AI model that supports 200+ languages, covering low-resource languages.
- Community-driven initiatives like Masakhane (for African languages) and AI4Bharat (for Indian languages), which are among the prominent open-source initiatives in creating improved translation models for less-resourced language groups.
Despite such developments, the presentation underscored the following problems:
- Insufficient training data: Many low-resource languages lack several large bilingual corpora for AI models to train on.
- One-way translation bias: The majority of the models are targeting translation from low-resource to high-resource languages, and not vice versa (which is essential for public governance).
- Inadequate domain accuracy: AI models are unable to translate legal, administrative, and technical documents, rendering them untrustworthy for governmental utilization.
Recent Advances in AI and Their Implications
The talk then delved into key AI advancements that hold promise in alleviating these challenges:
- Multilingual and Zero-Shot Learning – Translations into languages that the models have never trained on explicitly are now achievable through the utilization of knowledge from closely related languages.
- Knowledge Distillation – Techniques that allow large AI models to “teach” smaller models, allowing translation to become more effective for low-resource languages without gigantic datasets.
- Retrieval-Augmented Translation – Translations powered by artificial intelligence using contextual cues from previously existing bilingual texts, making complex documents more accurate.
- Zero-Resource Machine Translation – AI models that learn to translate from monolingual text only, enabling the possibility of building machine translation for languages where there are no parallel corpora.
Case Studies: Frisian and Asian Low-Resource Languages
To illustrate the practical effect of these developments, the presentation showed case studies of low-resource languages in government environments:
- Frisian (Germany and Netherlands) – Despite the close relation of Frisian to Dutch, its absence of digital corpora causes AI translation to fare poorly in government documents, necessitating novel solutions such as exploiting linguistic proximity to Dutch for enhancing machine translation.
- Lao & Khmer (Southeast Asia) – Having been left out of AI translation technology before, the languages were launched in Google Translate in recent years with zero-resource methods, revealing the possibility of AI bringing underserved languages into digital government.
- Sinhala and Nepali (South Asia) – By incorporating Sinhala and Nepali into a multilingual model with related languages, these languages with millions of speakers, but with very few digital resources, got a translation model.
Policy Implications and the Role of Governance
The presentation concluded with policy suggestions on how AI translation can be incorporated into public administration:
- Governments need to invest in open linguistic data – National and regional governments have to digitize official translations and contribute to public datasets.
- Public-private partnerships need to be increased – AI firms, universities, and governments need to work together on customized translation solutions for minority languages.
- AI translation needs to be tested for fairness – Policy-makers ought to screen AI bias in translation systems and be certain that machine-generated translations truthfully express linguistic diversity.
- Multilingual AI must be a digital public good – Governments and international organizations need to lead AI translation as part of the Global Digital Compact’s pledge for digital inclusivity.
Final Thoughts
In a world where AI is fast revolutionizing human communication, parity of linguistic representation in the virtual sphere is not only a technical issue—it is an issue of social justice and democratic participation. If AI translation as a bridge is to be, then the bridge has to be two-way. All languages—not only the hegemonic ones—need to be represented in digital governance.
As we march into an increasingly AI-powered future, we need to make sure that all languages have a say in the digital future.