UN partners with Google to open global data to AI agents

The United Nations announced Thursday a partnership with Google to build UN System Data Commons, a platform that will let language models and AI agents retrieve the organization's official statistics through natural-language queries. The system replaces the legacy UNData portal, which required manual searches through a traditional database interface, and adds support for the Model Context Protocol (MCP), the standard that allows AI to connect directly to external data sources.

The move follows a measurable problem. A UNICEF benchmark of six large models — OpenAI's GPT-4o and GPT-4o-mini, Anthropic's Claude Sonnet 4.5 and Haiku 4.5, and Google's Gemini 2.5 Flash and Gemini 2.0 Flash — across more than 133 thousand responses to questions on global development indicators yielded an average accuracy of just 21.2%. According to João Pedro Azevedo, UNICEF's chief statistician, roughly three out of five answers failed to provide a usable figure at all, often because the models "evaded" a direct response. When the same question set was run again two days later, models that supplied a number in both runs returned the same value only about half the time.

The research, still a UNICEF working paper that has not undergone peer review, is slated for publication with open code, data and methodology. Meanwhile, UNICEF's data site, which receives more than 6 million visits a month, is seeing a surge in traffic from generative AI agents: referrals from ChatGPT answers jumped 67% year-over-year between January and September, reaching 6.4% of all sessions. The estimate is that AI-powered assistants overall now account for roughly one-tenth of visits.

26 UN entities have committed to join Data Commons, and data from nearly 20 of them is already available at launch