UN teams with Google to make its global statistics readable by AI agents

The United Nations said on Thursday that it is working with Google to rebuild its collection of global statistics so AI systems can find and use it reliably. The new platform, called the UN System Data Commons, is built on Google's open-source Data Commons and lets users search statistics from across UN agencies with natural-language queries, replacing the older UNData portal where people largely had to browse a traditional database interface.

The detail that matters most to developers is support for the Model Context Protocol, the standard that lets AI systems connect directly to external data sources. Instead of relying on training data that may be stale, an agent can pull official UN statistics on demand, and the platform tracks where each number came from so an AI-generated answer can be traced back to its original source.

The UN's push follows evidence that models handle this kind of data poorly. João Pedro Azevedo, UNICEF's chief statistician, told reporters that a UNICEF benchmark of six large language models across more than 133,000 responses on global development indicators produced an average accuracy score of just 21.2%. The test covered 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.

Consistency was as much a problem as accuracy. About three in five responses did not provide a usable number at all, often because the models hedged their answers, Azevedo said. When the same questions were run again on the same model versions roughly two days later, models that produced a number both times returned the identical value only about half the time. The study is a UNICEF working paper being prepared for journal submission and has not been peer-reviewed; the organisation plans to release its methodology, code and data alongside the paper.

Demand from AI assistants is already visible in the traffic. UNICEF's data website receives more than six million visits a month, Azevedo said, and referrals from links inside ChatGPT answers rose 67% year over year between January 1 and September 14, accounting for 6.4% of all sessions this year; AI assistants overall now drive roughly one in ten visits.

On coverage, the UN said 26 of its entities have committed to the Data Commons, with data from nearly 20 available at launch, and it aims to bring 80% of the UN system's statistical datasets onto the platform by 2027. Google.org provided $2 million in capacity-building funding and technical support for the core infrastructure. Prem Ramaswamy, who leads Google's Data Commons team, said the system is hosted on a UN-governed instance and is intended to be maintained, operated and scaled independently by the UN.

A demonstration showed what that unlocks: connected through MCP, an AI system asked about the impact of the U.S. President's Emergency Plan for AIDS Relief in Africa pulled together UN statistics on HIV infections, AIDS mortality and life expectancy and produced an infographic, without a user manually assembling the underlying datasets.

Ramaswamy added a caveat that applies to every data-to-agent project: authoritative inputs do not guarantee authoritative conclusions. "Because models can misinterpret nuance, a human should always review the outputs before citing or publishing them," he said. For the UN, the platform is both a fix for AI accuracy and a play for visibility — if assistants become the default way people reach development data, being machine-readable with traceable sourcing is how the institution stays in the answer.