Center for Democracy & Technology

The Global Dialogue has already committed to ensuring linguistic diversity is part of the global AI governance conversation. The Dialogue must ensure conversations move beyond documenting the "digital language gap" and move towards filling current technical and research gaps. Tangibly, this requires: - Fostering an ecosystem of open systems to enable developers to fine tune and build language-specific tools. - Establishing multistakeholder channels to incorporate language and domain experts, including in the application of NetMundial +10 Sao Paulo Multistakeholder Guidelines - Creating a repository of independently-created multilingual evaluations or urging existing AISIs to adopt these evaluations to incentivize their use. - Directing funding towards language-specific research networks Read more at: https://cdt.org/insights/cdt-and-cornell-global-ai-initiative-call-for-meaningful-advancements-and-investment-into-linguistic-diversity-at-the-first-un-global-dialogue-on-ai-governance/

• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Open-source software, open data and open AI models
• Transparency, accountability, and human oversight

Currently, developers and deployers trying to build systems that work equally well across the world's 7000 languages face a practical constraint: There is very little high quality data in many of the world's major languages. While there are now more public AI training datasets representing these historically under-represented languages, an audit conducted by the Data Provenance Initiative in 2024 found that the relative representation of different languages and regions in the data used by large AI models has not changed since 2013. Beyond data availability issues, the representativeness of existing datasets in many language families is also an issue. Few datasets represent the specific terms and concepts people use in the domains where AI technologies may be deployed, such as when seeking healthcare information. This dearth in availability of high quality representative data limits model performance and capabilities in non-English languages and contexts. In an effort to bridge this gap, AI developers rely on a range of alternative sources of data, but much of it is imperfect. Synthetic data (that is, data generated using AI tools or English-language data translated using AI translation tools), data provided by government actors, and data in other semantically-similar languages are often used to train systems, but all of these tactics raise both practical and human rights concerns. As the International AI Safety Report outlines, advancements in AI capabilities remain jagged, particularly in languages other than English.

Directing funding towards language-specific research networks AI research groups that lead development of tools and paradigms for building AI systems in the world's languages such as Masakhane, SEA-LION, IndoNLP, AmericasNLP, and ARBML often have deep expertise about where the largest gaps are in their language's specific research but are sorely lacking the funds necessary to address them. The Global Dialogue can convene philanthropy, academia, and independent research groups to ensure a level playing field when it comes to research and development of multilingual tools. What's at stake when systems do not work equally well in languages other than English? Users may not be able to equally access and benefit from advancements in digital technology, creating what some experts call a "digital linguistic divide" and hampering progress towards SDG goals, including SDG 10 on "Reduced inequalities" and SDG 16 on "peace, justice, and strong institutions". The latter is relevant here as AI systems are increasingly deployed by governments and other institutions across a range of domains from healthcare settings to being used to determine who receives public services. These systems must serve all people regardless of the language they speak.

The shortcomings of these and other approaches are scarcely documented and mitigated because AI developers and deployers rarely conduct robust testing and evaluations in languages other than English or do so with imprecise instruments. CDT has outlined shortcomings of automated benchmarking tools broadly and how they should be improved to be more robust and pursue depth rather than simply breadth. In languages other than English, these automated tools used to judge outputs or evaluate systems are often more prone to failure as many languages are scarcely represented in available evaluation tools. Or, evaluation tools are not valid or suitable for the domain in which the AI system will be used. This also limits model developers and deployers from testing how well multilingual safety guardrails work in non-English contexts, making AI governance levers proposed in this and other arenas insufficient if they are easily jailbroken or circumvented in languages other than English.

As stated above, AI developers and deployers don't always have access to high quality evaluations in all of the world's languages. This can be particularly important when usage of a system evolves over time; for example, if a deployer sees an increase in use by Brazilian Portuguese speakers but does not easily have access to expertise in that language and context to test if model safeguards are proving effective. Having independent community-created evaluations to test systems' performance and safety guardrails in a specific language can identify gaps and prioritize collaboration and development of needed resources . Right now, evaluations developed by native language speakers aren't easily available or accessible and searching for the right one may be time consuming. An independent body can equip developers and deployers with appropriate evaluations and set up criteria to enable developers and deployers to choose the right evaluation. ML Commons has sought to fill this role to some degree but lacks the resources to serve as a comprehensive central repository; nonetheless, like other researchers, it has also sought to improve adoption of high quality evaluation tools by both creating a multilingual benchmarking suite and developing guidance to deployers of evaluations on how to best use it. AI red-teaming expert Roya Pakzad has also developed a platform for multilingual evaluations to be used by non-technical deployers and procurers of systems to increase informed selection of AI tools, which is another function a central repository can play.

Due to the dearth of existing resources in many non-English languages (datasets, evaluations, and more), participation of native language speakers, subject matter experts, and domain experts is critical to developing multilingual AI systems. One pilot program led by Microsoft Research India has shown that on-the-ground healthcare workers are best positioned to shape the data collection and annotation processes required to test AI systems used in healthcare settings in Indian languages towards more useful and effective performance of their systems. These channels for participation are essential when it comes to AI systems used in high risk settings as Cornell research has shown. The Global Dialogue can elevate these types of examples, identify best practices within them through the application of the NetMundial +10 Sao Paulo Multistakeholder Guidelines, and encourage AI companies to develop channels and use multistakeholder-developed best practices to incorporate those with the most expertise and linguistic fluency into the process when they develop and deploy AI systems.

Global Majority perspectives and global research communities are at the helm of improving language expertise, they must be included.

Virtual gatherings, gatherings in different time zones, and contributions on a tentative agenda can help ensure the first AI Dialogue advances a global agenda.

Please see our full vision here, co-submitted by Aliya Bhatia at the Center for Democracy & Technology, and Aditya Vashistha, Cornell Global AI Initiative. https://cdt.org/insights/cdt-and-cornell-global-ai-initiative-call-for-meaningful-advancements-and-investment-into-linguistic-diversity-at-the-first-un-global-dialogue-on-ai-governance/