UN and Google launch AI-ready Data Commons for global statistics

What launched

The United Nations has launched the UN System Data Commons, a platform built on Google's open-source Data Commons project that consolidates statistics from across UN agencies into a single, searchable knowledge graph. The partnership was announced Thursday, according to TechCrunch.

The platform replaces UNData, the previous portal where users navigated a conventional database interface. The new system supports natural-language search, so questions such as how life expectancy has shifted across regions, or how many people gained electricity access in the past decade, return relevant figures and interactive visualizations, Google's announcement explains. An Explore tab lets users filter by location and themes like health and education.

Data from nearly 20 UN entities is available at launch, with 26 entities committed overall. The UN aims to have 80% of the system's statistical datasets on the platform by 2027, a target cited by both TechCrunch and Google. Google.org provided $2 million in capacity-building funding plus technical support, and the platform runs on a UN-governed instance intended to eventually be maintained and scaled by the UN itself, Prem Ramaswami, who leads Google's Data Commons team, told TechCrunch.

How AI agents connect

Beyond the human-facing interface, the platform supports the Model Context Protocol (MCP), the open standard that lets AI systems connect directly to external data sources. Google added MCP support to Data Commons last year, and the UN instance extends that capability: agents can autonomously fetch official figures, combine indicators across domains, and package results into charts, infographics, or draft reports without anyone manually assembling spreadsheets.

In one demonstration reported by TechCrunch, Google asked an AI system to assess the impact of the U.S. President's Emergency Plan for AIDS Relief in Africa. The system identified UN statistics on HIV infections, AIDS mortality, and life expectancy, then produced an infographic from them.

Each statistic carries provenance, so answers surfaced by an AI can be traced back to the original UN source. Ramaswami cautioned that grounded data does not guarantee grounded conclusions: because models can misinterpret nuance, people should review outputs before citing or publishing them.

The accuracy problem it targets

The launch responds to how people now look for numbers. A UNICEF benchmark described to reporters by chief statistician João Pedro Azevedo tested six large language models on more than 133,000 questions about global development indicators and produced an average accuracy score of just 21.2%. The models tested were 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.

About three in five responses gave no usable number at all, often because the models hedged, Azevedo said. When the same questions were re-run days later on the same model versions, models that answered twice returned the identical figure only about half the time. The study is a working paper being prepared for journal submission and has not been peer-reviewed, though UNICEF plans to release its methodology, code, and data.

AI assistants are also becoming a meaningful traffic source for UN data. UNICEF's data site, which draws more than 6 million visits a month, saw clicks from ChatGPT answers rise 67% year over year between January 1 and September 14; such referrals made up 6.4% of sessions this year, and the agency estimates AI assistants drive roughly one in ten visits overall.

Shantanu Mukherjee, acting director of the UN Statistics Division, told reporters the platform connects across more UN agencies than previously possible while making the data, in his words, AI-ready.

Why it matters

Official statistics that models cannot reliably retrieve are effectively invisible to a growing share of the public. If around a tenth of visits to a major UN data site already originate from AI assistants, the real interface for official numbers is increasingly an agent rather than a browser. MCP-based access with per-figure provenance gives model builders a way to ground answers in validated UN data instead of stale training corpora, and the UNICEF benchmark's 21.2% accuracy score shows how wide the gap is today. The open-source foundation, UN-governed hosting, and Google's train-the-trainer handover also offer a replicable template for other public institutions whose data sits in incompatible silos. The caveat remains: verified inputs do not make generated analysis authoritative, and human review stays part of the workflow.