The United Nations launched a new platform on Thursday. It aims to gradually consolidate statistics from all its agencies and prepare them so that AI-powered programs can work with them as well. It is called the UN System Data Commons and replaces the older UNData portal, where users had to search for data by navigating through a traditional database. The new version works differently. You can ask a question in everyday language, such as how the number of people with access to electricity has changed over the past ten years, and receive an answer complete with a chart.
Reason for the Changes
The catalyst was a test conducted by UNICEF. Its chief statistician, João Pedro Azevedo, told reporters how they had six large language models answer questions about global development indicators. The test generated more than 133,000 answers, with an average success rate of roughly one fifth. Models from OpenAI, Anthropic, and Google were tested.
Even more interesting is how exactly the models failed. About three out of five answers contained no usable figure at all, because the model avoided giving a direct answer and opted for a general statement instead. When the researchers asked the same versions of the models the same questions again two days later, they received the same figure as before in only half of the cases.
Meanwhile, people are increasingly accessing UN data through chatbots. UNICEF's data website receives more than six million visits per month, and traffic from links in ChatGPT answers has jumped by two thirds year over year this year. The agency estimates that roughly one in ten visitors now comes from an AI assistant.
Why They Chose Google
Google has operated the Data Commons project since 2018. It is an open platform that combines public datasets from various sources into a standardized format, allowing them to be compared with one another. Moreover, the UN and Google were not starting from scratch, as they have been collaborating since 2023, when a smaller version focused on the Sustainable Development Goals was created.
Another reason is that last year Google added support for the MCP protocol to Data Commons. This is an agreed-upon method that allows AI programs to connect directly to external data sources. Without it, a model constructs its answer from what it remembers from training. With it, the model can look up a specific figure in a specific table and know where it came from. For every data point, the UN platform maintains a trail back to its source, making it possible to trace what the AI based its answer on.
The data itself is not stored by Google. The system runs on a UN-managed instance, and the goal is for the organization to operate and expand it independently over time. Prem Ramaswamy, who leads Data Commons at Google, describes it as a train-the-trainer approach, in which the UN team learns to operate the platform on its own.
What It Will Be Able to Do
In addition to looking up individual figures, Google demonstrated how AI connected through MCP can combine several indicators and immediately use them to create a dashboard, charts, or a written analysis. In one demonstration, the system was tasked with determining the impact of a U.S. aid program aimed at combating AIDS in Africa. It found statistics on new infections, mortality, and life expectancy and turned them into an infographic. An analyst would take several days to do the same work manually, as they would have to download and standardize data from several different agencies.
Source: techcrunch.com
