Center for Global Development

1) It should be a genuinely Global dialogue that includes stakeholders from all parts of the UN community. 2) It should cover difficult issues as well as conventional wisdom ones: examples are AI sovereignty (especially the implications for LMICs), reliability on a few large firms, weaponisation of the technology and of access to it as a tool of power politics. 3) It should lead to tangible initiatives that help countries and societies to coordinate their efforts, find common ground, and join forces.

AI capacity-building;Social, economic, ethical, cultural, linguistic and technical implications of AI;Safe, secure and trustworthy AI

The Center for Global Development's AI Initiative works on AI in low- and middle-income country (LMIC) governments, where deployment is accelerating but governance capacity and local evidence is pretty thin on the ground. Safe, secure and trustworthy AI: Current safety discourse centres frontier-model risks salient in high-income settings. Yet LMIC governments are already procuring AI for social protection targeting, agricultural advisory, health triage, and tax administration. The relevant failures, which might include biased eligibility decisions (I'm a researcher working on targeting, so this is close to my heart), unreliable advice in low-resource languages, or vendor lock-in, receive far less attention. Urgent action is needed on evaluation standards, audit mechanisms, and procurement guardrails that fit capacity-constrained public-sector buyers. AI capacity-building: Without technical capacity within LMIC governments to specify, evaluate, and oversee AI systems, deployments default to vendor-led design with weak public benefit and accountability. Capacity-building should target the functions that determine whether AI in government works: procurement, evaluation across the full impact chain (model performance, user engagement, behavioural change, outcomes), and regulatory oversight. Social, economic, ethical, cultural, linguistic and technical implications: Foundation models are predominantly trained on high-resource languages and high-income contexts. Performance gaps in low-resource languages, cultural mismatches in interfaces, and labour-market effects in services-dependent economies are first-order issues for billions of users yet remain underexamined. Sustained investment is needed in evidence on AI's distributional consequences in LMICs and in datasets, benchmarks, and tools fit for non-Western contexts. Across all three, I'd say that international AI governance must include LMIC perspectives and address the practical deployment realities that will shape AI's welfare impact for most of the world's population.

For me personally, coming from a think tank, strategic issues such as AI sovereignty would be good to discuss in a pointed fashion, not just vague diplomatic discourse.

Working in the UK office of a global think tank, our perspective is not locally tied or related to a particular sector. Personally, I've been working most on social protection in the past years, and here the concerns are about the oversight of activities in the public sector, which should include both elected representative and civil society more than has been the case: AI is currently treated as a technocratic topic that only experts can understand. Opening it up to wider scrutiny is especially important in this sector that is no only very large, but also crucial for beneficiaries.