Product1 publisher3 min readPublished
The UN rebuilds its statistics portal on Google's Data Commons with an MCP endpoint for agents
UNICEF's chief statistician says six frontier models averaged 21.2% accuracy across more than 133,000 questions about development indicators, and the UN's answer is a portal that hands agents the numbers with their sources attached.
The Product Desk · Product desk
What happened
- The United Nations said on Thursday it is working with Google to make its global statistics easier for AI systems to access and use, through a new portal called the UN System Data Commons.
- The platform is built on Google's open source Data Commons software, takes natural-language queries across UN agencies, and replaces the UNData portal and its database-style browse and search interface.
- UNICEF chief statistician Joao Pedro Azevedo told reporters that a benchmark of six large language models across more than 133,000 responses on global development indicators averaged 21.2% accuracy.
- Twenty-six UN entities have committed to the platform, data from nearly 20 was available at launch, and the UN wants 80% of the system's statistical datasets on it by 2027.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Any organisation that publishes numbers other people cite now has a choice with a visible default: stand up a machine-readable endpoint, or let assistants answer questions about your indicators from whatever they absorbed in training.
- exposure UNICEF's most popular data site now gets about one visit in ten from AI assistants, so how a small number of vendors decide to cite sources moves a share of the agency's audience that no site redesign can win back.
- constraint Retrieval fixes the models that would not commit to a figure at all; it does not fix a model that returns a different figure two days later, so anything an agent composes from UN data still needs a person checking the numbers against source.
- cost Google.org's $2 million covers core infrastructure and training, and the UN carries the rest: migrating datasets toward its 80% target and eventually operating the instance itself.
About three in five times, a question about a global development indicator came back from these models with no usable number in it, usually because the model hedged, Azevedo said [6]. If a hedge scores zero, the two in five answers that did contain a number were right roughly half the time: 21.2 divided by 40 is 53 percent [1]. One of those failures is a retrieval problem, and an endpoint addresses it. The other showed up when the team put the same questions to the same model versions about two days later. Among answers that carried a number both times, the two numbers matched only about half the time [7].
The UN instance answers over the Model Context Protocol, which Google added to Data Commons last year so agents could query statistics and their sources directly [16]. Each statistic keeps its provenance, so a figure an AI system retrieved can be traced back to the original UN source [17]. The briefing as reported did not include an accuracy score for the same benchmark questions answered with that endpoint attached. UNICEF's study is a working paper being prepared for journal submission and has not been peer reviewed; the agency said it will publish the methodology, code and data alongside it [8].
The demand is visible in UNICEF's own logs. Its data site takes more than 6 million visits a month and is among the agency's most popular [9]. Clicks from links in ChatGPT answers rose 67% year over year between January 1 and September 14 and made up 6.4% of all sessions this year, while UNICEF estimates AI assistants together account for about one visit in 10 [10]. On those numbers ChatGPT alone is roughly two thirds of the assistant traffic [3], and 6.4% of 6 million works out at about 380,000 sessions a month [4].
"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," said Shantanu Mukherjee, acting director of the UN Statistics Division [12]. Prem Ramaswami, who leads Google's Data Commons team, told TechCrunch the platform runs on a UN-governed instance and is meant to be maintained, operated and scaled by the UN itself [14]. Ramaswami said the rollout has taken a train-the-trainer approach throughout, and that the UN system team has already ramped up quickly [15].
People get something here as well: the same platform takes natural-language queries, where UNData mostly made users browse and search a database interface [2]. The part that goes past retrieval is composition. Google showed an AI system pulling multiple indicators through MCP and assembling dashboards, charts and written analysis without anyone combining the datasets by hand [18]. Asked about the impact of the U.S. President's Emergency Plan for AIDS Relief in Africa, the system picked statistics on HIV infections, AIDS mortality and life expectancy and produced an infographic [19]. As TechCrunch noted, giving an AI system authoritative data does not necessarily make its conclusions authoritative [20].
A statistics team can run the same two-part check on its own indicators in an afternoon: one of its headline numbers put to a model with no tools, then the same question two days later. Blanks and hedges are a publishing problem, and an MCP endpoint is the fix now on offer. If the number changes between the two runs, that belongs to the model. No endpoint corrects it.
What to watch
- Publication of the UNICEF working paper with its methodology, code and data, and whether peer review holds the 21.2% average.
- Whether the roughly six committed UN entities missing at launch load their datasets, and how close the platform gets to 80% coverage by 2027.
- A rerun of the same benchmark questions with the UN endpoint attached, which would show whether retrieval moves the score.