The United Nations announced Thursday that it is working with Google to make its vast collection of global statistics more accessible and usable by AI systems.
The new system, called UN System Data Commons, is built on Google’s open source Data Commons platform and allows people to search statistics across UN agencies using natural language queries. It replaces the existing UNData portal, which required users to browse and search statistics primarily through a traditional database interface. The new platform also supports Model Context Protocol (MCP), a standard that allows AI systems to connect directly to external data sources.
Users are increasingly relying on AI tools to get answers, but many systems still struggle to reliably display reliable data. João Pedro Azevedo, UNICEF’s chief statistician, told reporters in a virtual briefing that the average accuracy score of a benchmark of six large language models based on more than 133,000 responses to questions on global development indicators was just 21.2%.
Azevedo told TechCrunch that the tests included 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.
Three out of five responses did not provide any usable numbers, Azevedo said. This is often because the model hedged the answer. However, when we ran the same question again with the same model version about two days later, the model that both provided numbers returned the same number only about half of the time.
This study is a UNICEF working paper, is being prepared for submission to a journal, and has not yet been peer-reviewed. The organization said it plans to publish its methodology, code and data at the same time as the paper.
UNICEF has seen a surge in traffic from its generative AI assistants to its data website this year. The website receives over 6 million visitors per month and is one of UNICEF’s most popular websites. Azevedo told TechCrunch that the number of visits from users who clicked on a link in a ChatGPT answer to the site increased 67% year-over-year between January 1 and September 14. While such referrals accounted for 6.4% of all sessions this year, UNICEF estimates that AI assistants overall now account for around 1 in 10 visits.
The United Nations said 26 of its agencies are participating in the Data Commons, and data from nearly 20 agencies will be available at launch. Furthermore, we aim to have 80% of the United Nations system’s statistical datasets on the platform by 2027.
“We are making an order of magnitude advance in scale, scope and flexibility, bringing together so many agencies across the UN system for the first time,” said Shantanu Mukherjee, acting director-general of the UN Statistics Division. “And[we]are also taking advantage of this moment to make data AI-enabled.”
Google.org provided $2 million in capacity-building funds and technical support to establish the platform’s core infrastructure. Prem Ramaswamy, who leads Google’s Data Commons team, told TechCrunch that the system is hosted on a UN-controlled instance and will eventually be independently maintained, operated and scaled by the UN.
“We have adopted a ‘train the trainer’ approach throughout the deployment and have already seen the rapid strengthening of UN system teams,” Ramaswamy said.
Google launched Data Commons in 2018 in an effort to organize public datasets from disparate sources into a common framework. Last year, we added support for MCP, allowing AI agents to directly query the data commons for statistics and their sources.
The UN platform also tracks the origin of each statistic, allowing people to trace data captured by AI systems back to the original UN source. Azevedo told reporters this is important because more and more people rely on AI tools to find and interpret information.
Along with enabling AI agents to retrieve individual statistics, Google demonstrated how an AI system connected to UN data via MCP can be used to pull together multiple metrics to generate dashboards, graphs, and written analysis without users having to manually find and combine the underlying datasets.
In one demonstration, Google asked its AI system to examine the impact of the U.S. president’s emergency plan for AIDS relief in Africa. The system identified relevant United Nations statistics on measures such as HIV infection, AIDS mortality, and life expectancy, and used them to create infographics.
However, feeding an AI system reliable data does not guarantee that its conclusions will be reliable. “Models can misunderstand nuances, so a human should always review the output before quoting or publishing it,” Ramaswamy said.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
