The United Nations announced Thursday that it's partnering with Google to rebuild its worldwide data portal so artificial intelligence systems can tap directly into authoritative global statistics. The new UN System Data Commons replaces the older UNData portal and runs on Google's open-source platform, letting users query UN statistics in plain language instead of navigating traditional database menus. The initiative comes as a UNICEF study of six major language models found they answered questions about development indicators correctly just 21.2% of the time across more than 133,000 responses.

The UNICEF benchmark, which tested 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, revealed that roughly three out of five responses failed to provide any usable figure, often because the models hedged their answers, according to João Pedro Azevedo, UNICEF's chief statistician. When the identical questions were posed again approximately two days later, models that supplied a number in both rounds returned the same figure only about half the time. Meanwhile, UNICEF's data website has seen visits from users clicking ChatGPT links jump 67% year-over-year between January 1 and September 14, with such referrals now representing 6.4% of all sessions and AI assistants overall accounting for roughly one in 10 visits to a site that draws more than 6 million monthly visitors.

At launch, data from nearly 20 of the 26 UN entities that have committed to the platform is already available, and the organization aims to migrate 80% of the UN system's statistical datasets onto the system by 2027. Shantanu Mukherjee, acting director of the UN Statistics Division, said the initiative represents advances "orders of magnitude" larger in scale, scope, and flexibility by connecting agencies across the UN system for the first time while making the data AI-ready. Google.org contributed $2 million in funding and technical support to build the platform's core infrastructure, and Prem Ramaswami, who leads Google's Data Commons team, noted the system is hosted on a UN-governed instance designed to eventually be maintained and scaled independently by the UN.

The platform supports the Model Context Protocol, a standard allowing AI systems to connect directly to external data sources and trace each statistic back to its original UN source—a feature Azevedo described as important as more people depend on AI tools to locate and interpret information. In demonstrations, Google showed how an AI system linked to UN data through the protocol could assemble multiple indicators to generate dashboards, charts, and written analysis without users manually finding and combining datasets; one example involved asking the system to assess the impact of the U.S. President's Emergency Plan for AIDS Relief in Africa, prompting it to identify relevant statistics on HIV infections, AIDS mortality, and life expectancy and produce an infographic. However, Ramaswami cautioned that supplying an AI system with authoritative data doesn't automatically make its conclusions authoritative, stating that "a human should always review the outputs before citing or publishing them" because models can misinterpret nuance. The organization's working paper on the benchmark study is being prepared for journal submission and hasn't yet undergone peer review, with UNICEF planning to release its methodology, code, and data alongside the paper. For organizations trying to balance automation with accuracy, the partnership suggests that infrastructure alone won't solve hallucination problems if the reasoning layer remains unchecked. The real test will be whether decision-makers treat AI-generated insights as starting points rather than final answers.