The United Nations announced on Thursday that it is partnering with Google to rebuild how the world's authoritative statistics reach AI systems, after internal testing showed leading large language models answer questions about global development indicators correctly just 21.2% of the time1.
The new platform, called UN System Data Commons, is built on Google's open-source Data Commons infrastructure and replaces the existing UNData portal. It lets users query statistics from across UN agencies using natural language and supports Model Context Protocol, the standard that allows AI agents to connect directly to external data sources2.
How badly models fail today
A UNICEF benchmark tested six models across more than 133,000 responses to questions about global development indicators. The models tested were OpenAI's GPT-4o and GPT-4o mini, Anthropic's Claude Sonnet 4.5 and Claude Haiku 4.5, and Google's Gemini 2.5 Flash and Gemini 2.0 Flash. The average accuracy was 21.2%, according to João Pedro Azevedo, UNICEF's chief statistician.
About three in five responses did not provide a usable number at all, often because the models hedged their answers. When the same questions were repeated roughly two days later on the same model versions, models that did return a number both times gave the identical figure only about half the time. The study is a working paper being prepared for journal submission and has not yet been peer-reviewed. UNICEF plans to release its methodology, code, and data alongside the paper.
Platform scope and funding
Twenty-six UN entities have committed to the Data Commons, and data from nearly 20 are available at launch. The UN aims to bring 80% of its system's statistical datasets onto the platform by 2027. Google.org provided $2 million in capacity-building funding and technical support to establish the platform's core infrastructure. The system is hosted on a UN-governed instance and is intended to eventually be maintained, operated, and scaled independently by the UN.
Google launched Data Commons in 2018. MCP support was added to Data Commons last year, enabling AI agents to query statistical databases directly and verify provenance. The UN platform tracks where each statistic originates so users can trace AI-retrieved data back to the original UN source.
Demand signal and demonstrations
Visits to UNICEF's data site from users clicking links in ChatGPT answers rose 67% year-over-year between January 1 and September 14, according to Azevedo. Such referrals accounted for 6.4% of all sessions this year, and UNICEF estimates that AI assistants overall now account for about one in 10 visits.
In a demonstration, Google showed that an AI system connected to UN data through MCP could pull together multiple indicators and generate dashboards, charts, and written analysis. One example had the system analyze the impact of PEPFAR in Africa, using UN statistics on HIV infections, AIDS mortality, and life expectancy to produce an infographic. Prem Ramaswami said a human should always review AI outputs before citing or publishing them, because a model may misunderstand context or fail to account for important nuances even when it receives authoritative data.
ANALYSIS The 21.2% accuracy figure and the inconsistency across repeat queries quantify a problem the Data Commons is designed to address structurally: rather than relying on models to recall statistics from training data, MCP lets agents retrieve verified figures at query time. The platform's provenance-tracking layer adds a verification step that raw model outputs lack.