The United Nations has started working with Google to make comprehensive global statistics held across its different organizations more easily accessible to both users and artificial intelligence systems. announced on Thursday UN System Data Commonsis based on Google’s open source Data Commons infrastructure and allows statistics from UN agencies to be searched with natural language questions. The new system replaces the existing UNData portal, which relies on more traditional database searching and navigation methods to access statistics. One of the striking aspects of the platform is that it offers support for Model Context Protocol, also known as MCP, which allows artificial intelligence systems to directly connect to external data sources. Thus, instead of creating a statistics portal that is only visited by people, the UN aims to transform its data into a structure that artificial intelligence agents can directly query.
The reason behind this change is that users are increasingly turning to productive artificial intelligence tools to obtain information. However, large language models still have serious shortcomings in accurately finding and conveying reliable and official statistics. According to data shared by UNICEF Chief Statistician João Pedro Azevedo, in a test conducted by the agency, the six major language models achieved an average accuracy of only 21.2 percent across more than 133 thousand responses on global development indicators. OpenAI GPT-4o and GPT-4o-mini, Anthropic Claude Sonnet 4.5 and Haiku 4.5, and Google Gemini 2.5 Flash and Gemini 2.0 Flash models were evaluated in the study. It was stated that approximately three out of every five answers did not contain a usable numerical value, and in most cases the models used conservative expressions rather than giving precise data.
The test also revealed results examining whether large language models provide consistent answers to the same question at different times. When the same questions were posed again to the same model versions about two days later, the same figure was reached with only about half of the models giving a numerical answer on both attempts. Therefore, the problem is not limited to finding the right statistics; The repeatability of the response is also a separate issue. UNICEF’s research is not currently a definitive peer-reviewed academic study. The institution plans to publish the methodology, codes and data used together with the working article prepared to be sent to the journal.
UN data will be connected to artificial intelligence agents via MCP
At this point, UN System Data Commons’ MCP support aims to provide direct access to the data source. Google introduced Data Commons in 2018 to organize public datasets from different sources under a common structure. With the addition of MCP support to the platform last year, it became possible for artificial intelligence agents to directly query the statistics in Data Commons and the sources of these statistics. The system to be used by the UN also keeps track of where each statistic comes from. Thus, the user can control which UN source the data brought by an artificial intelligence tool is based on.
Shantanu Mukherjee, Acting Director of the UN Statistics Division, says the new infrastructure is significantly more advanced than previous systems in terms of scale, scope and flexibility, and connects multiple UN agencies on this scale for the first time. According to the UN, 26 organizations have committed to participating in the UN System Data Commons, and data from approximately 20 of them will be available once the system is launched. The organization’s goal is to move 80 percent of the statistical datasets in the UN system to this platform by 2027. In addition, Google.org provided $2 million in capacity development financing and technical support to establish the basic infrastructure. Prem Ramaswami, head of the Google Data Commons team, states that the system is hosted in a separate environment managed by the UN, and that maintenance, operation and scaling processes are planned to be carried out entirely by the UN in the future.
Another indicator of the need for new infrastructure can be seen in UNICEF’s internet traffic. Traffic from productive AI assistants to the UNICEF data site, which receives more than 6 million monthly visits, has increased significantly this year. According to the information Azevedo gave to TechCrunch, between January 1 and September 14, visits to the UNICEF site from links in ChatGPT replies increased by 67 percent compared to the same period last year. These referrals will account for 6.4 percent of all sessions in 2026, while UNICEF estimates that different AI assistants drive approximately one-tenth of site visits in total. This table shows that the data prepared by official institutions is increasingly effective not only in how they are presented on their websites, but also in how they are found and transferred by artificial intelligence tools.
The MCP connection will not only be used to find individual statistics. In Google’s demonstration, an AI system connected to UN data was able to simultaneously find and combine different indicators and create dashboards, graphs and written analysis from them. In one example, the system was tasked with investigating the impact of the US President’s AIDS Emergency Plan on Africa. Artificial intelligence; He identified UN statistics on indicators such as HIV infections, AIDS-related deaths and life expectancy, then used them to create an infographic. This process, which would normally require manually finding and combining different data sets, can be carried out in a shorter time with a direct data connection.
However, having direct access to a reliable data source does not mean that the interpretations that artificial intelligence derives from this data are automatically reliable. Ramaswami also points out that models can misinterpret the nuances in the data and emphasizes that the artificial intelligence outputs to be published or cited should be checked by humans. UN System Data Commons therefore offers an infrastructure that improves the source and traceability of the data that models can refer to, rather than a system that eliminates artificial intelligence’s judgment errors. The main point that differs from the old UNData system for the UN emerges here; Statistics are no longer presented only as records that people can search for, but as data that software and artificial intelligence agents can query through a standard protocol. The extent to which this will be of practical benefit will depend on the models’ ability to use this data in the right context and the necessary human control of the generated outputs.