Civil society groups call for stronger focus on linguistic diversity in UN AI governance talks
The Center for Democracy & Technology and the Cornell Global AI Initiative have urged policymakers to address structural gaps in multilingual AI systems, warning that current approaches risk reinforcing global inequalities.
Civil society and academic actors are pushing for more concrete measures to address linguistic imbalances in artificial intelligence systems, as discussions on global AI governance continue at the United Nations.
In a joint submission to the UN’s Global Dialogue on AI Governance, the Center for Democracy & Technology and the Cornell Global AI Initiative argue that current AI development practices remain heavily skewed towards English and a limited set of well-resourced languages. Their comments build on earlier consultations and focus on how governance processes can better support multilingual and culturally diverse AI systems.
According to the submission, one of the central constraints is the lack of high-quality data across most of the world’s languages. While new datasets have emerged in recent years, their overall representation remains uneven, limiting the performance and reliability of AI systems in non-English contexts.
The groups also highlight risks linked to current workarounds, such as the use of synthetic or machine-translated data. These methods, while widely used, can introduce inaccuracies and raise concerns related to bias and accountability, particularly in sensitive domains like healthcare or public service delivery.
Beyond data gaps, the submission points to weaknesses in how AI systems are evaluated. Testing frameworks are often designed for English-language environments and may not capture failures in other linguistic or cultural contexts. This creates blind spots in assessing whether safety mechanisms and safeguards function effectively across different user groups.
To address these issues, the organisations propose a set of procedural measures within the Global Dialogue process. These include support for open-source models and datasets tailored to specific languages, stronger involvement of local language and domain experts in system development, and the creation of shared repositories for multilingual evaluation tools.
They also call for increased funding for regional research networks working on language technologies, arguing that such initiatives are often best positioned to identify local needs but lack sustained financial support.
The submission frames these gaps not only as technical challenges but as issues with broader implications for digital inclusion and human rights. If left unaddressed, unequal system performance across languages could limit access to digital services and deepen existing inequalities, particularly as AI systems are increasingly used in public sector decision-making.
