The United Nations and Google have launched the UN System Data Commons, built on top of Google's open-source Data Commons, replacing the long-running UNData portal. The new platform supports natural-language queries and the Model Context Protocol (MCP), letting AI systems pull data directly from the underlying sources and return results with traceable citations. Google is backing the effort with $2 million in funding through Google.org.
The push follows directly from a UNICEF study. The research tested six models on questions about global development indicators — 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 2.0 Flash — across more than 133,000 answers, with average accuracy at 21.2%. About 60% of answers gave no usable number, with models tending toward vague responses; asking the same question again two days later produced a matching figure only about half the time. The study is still an unpublished working paper that hasn't been submitted for peer review.
From browsing by humans to fetching by machines
UNData took the classic statistical-portal shape: click through by agency, by indicator, by year, then download a table. The new platform swaps that entry point for a single call. MCP acts as the connection layer here — models no longer scrape a cached webpage, they pull data straight from the source and bring back a citation path.
Shantanu Mukherjee, acting director of the UN Statistics Division, put it this way:
We are orders of magnitude more advanced in scale, scope, and flexibility.
Prem Ramaswami, who leads Google's Data Commons, described the rollout as a train-the-trainers approach, while adding a caveat: models can misread subtle nuance, so outputs should be reviewed by a person.
Twenty-six UN agencies have committed to joining so far, with close to 20 datasets available at launch. The goal is to bring 80% of the UN system's statistical datasets onto the platform by 2027.
The Chinese-language side is still blank
There's no equivalent access layer yet for China's public statistics. Data from the National Bureau of Statistics, individual ministries, and local statistical yearbooks is still mostly published as web tables and PDFs; if a model wants to use it, someone has to scrape and reassemble it by hand, and it's hard to trace where errors creep in. Ask a model today what China's urbanization rate was in 2024, and the number it returns most likely comes from some secondhand writeup rather than the original statistical source — the exact problem the UN is now trying to solve on its own side.
There's another metric worth watching: referral traffic. The study notes that ChatGPT-driven referrals grew 67% year-over-year between January and September 2026. Statistical agencies used to measure impact by portal visits; now there's a second thing to track — how often a model cites your data when answering a question.
What's still unanswered
Who pays for the platform's long-term upkeep, and where the budget comes from after the initial $2 million, wasn't addressed in the announcement. How granular the 26 agencies' data will get, how sensitive indicators will be handled, and how member states' concerns about data sovereignty will be managed also have no public plan yet. The 80%-by-2027 goal currently has just one target date, with no staged benchmarks for checking progress along the way.
Sources: TechCrunch, UN Statistics Division and Google Data Commons announcements, CocoLoop, UNICEF working paper; accuracy figures and agency counts cross-checked against the study and announcement.