The United Nations announced on Thursday that it is collaborating with Google to modernize its vast repository of global statistics, making them more accessible and usable for artificial intelligence systems.
Called the UN System Data Commons, the new infrastructure is built upon Google’s open-source Data Commons platform. It allows users to query statistics from across various UN agencies using natural language, replacing the legacy UNData portal which relied on traditional database interfaces. The platform also supports the Model Context Protocol (MCP), a standard enabling AI systems to connect directly to external data sources.
The initiative addresses a growing reliance on AI tools for information retrieval, despite current limitations in accuracy. During a virtual briefing, João Pedro Azevedo, chief statistician at UNICEF, revealed that a benchmark test of six large language models yielded an average accuracy score of just 21.2% when answering questions about global development indicators.
The study, which evaluated models including OpenAI’s GPT-4o, Anthropic’s Claude Sonnet 4.5, and Google’s Gemini variants across more than 133,000 responses, found that approximately 60% of answers did not provide a usable figure, often because the models hedged. Furthermore, when re-tested two days later, identical models produced consistent numbers only about half the time. UNICEF noted that the working paper is pending peer review but plans to release its methodology, code, and data alongside the publication.
Reflecting the shift in user behavior, UNICEF reported a 67% year-over-year increase in traffic from generative AI assistants to its data website between January 1 and September 14. The agency estimates that AI assistants now account for roughly one in 10 visits to its site, which receives over 6 million monthly sessions.
Shantanu Mukherjee, acting director of the UN Statistics Division, emphasized the scale of the upgrade. “We are orders of magnitude more advanced in scale, scope, and flexibility, connecting for the first time across so many agencies across the UN system,” Mukherjee said. “And we are taking this moment to also make our data AI-ready. “
Google.org contributed $2 million in capacity-building funding and technical support to establish the platform’s core infrastructure. Prem Ramaswami, who leads Google’s Data Commons team, stated that the system is hosted on a UN-governed instance and is designed for eventual independent operation by the UN. He highlighted a “train-the-trainer” approach that has allowed the UN system team to ramp up quickly.
Originally launched in 2018, Google’s Data Commons organizes public datasets into a common framework. Support for MCP was added last year to facilitate direct querying by AI agents. The UN platform retains provenance tracking, allowing users to trace AI-retrieved data back to its original source.
During a demonstration, Google showed how an AI system connected via MCP could synthesize multiple indicators to generate dashboards and written analysis. In one example, the system analyzed the impact of the U.S. President’s Emergency Plan for AIDS Relief in Africa by pulling data on HIV infections, mortality, and life expectancy to create an infographic.
Despite these advancements, officials cautioned against uncritical reliance on AI outputs. Ramaswami warned that because models can misinterpret nuance, human review is essential before citing or publishing any data generated by the system.
Love the direct MCP integration, but I’m skeptical about long-term UN independence. Google built the rails, who owns the train?
21 percent accuracy is terrifying. We really need that human review layer before anyone cites these AI stats anywhere.