Data Commons is now the engine behind the United Nations’ main statistics website. On 17 September 2026 the UN launched the UN System Data Commons, built on Google’s open-source platform, to replace the old UNdata portal at data.un.org. People can search it in plain language, and AI agents can query it directly through the Model Context Protocol (MCP).
The reason is blunt. A UNICEF test found leading AI models answered questions about global development statistics with an average accuracy of just 21.2%. If people are going to ask chatbots for numbers on child poverty or clean water, the UN wants those chatbots reading its own figures. This article explains what launched, how it works, what the benchmark found and what it means for anyone building with AI.
Table of contents
- What the UN Launched on 17 September
- Why the UN Needed an AI-Ready Data Commons
- How Data Commons Works
- How AI Agents Use Data Commons Through MCP
- UNdata Versus the UN System Data Commons
- AI Traffic Is Already Arriving
- From SDG Portal to System-Wide Data Commons
- What Could Go Wrong
- How to Connect an Agent in Practice
- What It Means for Businesses and Developers
- Data Commons FAQ
- References
What the UN Launched on 17 September
The launch was announced by the UN and Google and first reported by TechCrunch, which spoke to officials from both. Here is what is new.
A new front door at data.un.org
The UN System Data Commons replaces the UNdata portal, where users mostly browsed and searched a traditional database interface. The old site survives as a legacy archive. The new one opens with a natural-language search box, an Explore tab for filtering by place or theme, and a Blog section with ready-made reports.
Built on Google’s open-source platform
The system runs on Data Commons, which Google launched in 2018 to organise public datasets from different sources into one framework. Google calls the result “an AI-ready knowledge graph”. Every dataset, it says, “is validated with UN system statisticians and technical experts”.
Who pays and who runs it
Google.org provided $2 million in capacity-building funding and technical support for the core infrastructure. Prem Ramaswami, who leads the Data Commons team at Google, told TechCrunch the system is hosted on a UN-governed instance and is meant to be “maintained, operated, and scaled independently by the UN”. Google used what he called a “train-the-trainer” approach.
How big it is at launch
According to TechCrunch, 26 UN entities have committed to the platform and data from nearly 20 is available now. The target is to bring 80% of the UN system’s statistical datasets onto it by 2027. Shantanu Mukherjee, acting director of the UN Statistics Division, said the system is “orders of magnitude more advanced in scale, scope, and flexibility”.
Why the UN Needed an AI-Ready Data Commons
The case for the project rests on one piece of research. It is worth reading carefully, because it says as much about how AI fails as about how often.
The UNICEF benchmark
João Pedro Azevedo, UNICEF’s chief statistician, told reporters his team tested six leading models across more than 133,000 responses to questions about global development indicators. The average accuracy score was 21.2%.
Which models were tested
The six were 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. None of them is the newest model from its maker, which matters when reading the result.
Most answers had no usable number
About three in five responses did not give a usable number at all, often because the model hedged. That is a different failure from inventing a figure. It means the user leaves without an answer, or goes to find it somewhere else.
Same question, different answer
When the same questions were rerun on the same model versions about two days later, models that gave a number both times returned the identical number only about half the time. For statistics, inconsistency is as damaging as error.
“About 60%” is the reported “three in five”; “about 50%” is the reported “about half”. Bar widths are the percentages themselves.
Why models struggle with statistics
The result is less surprising than it sounds. Development statistics change every year, and most indicators exist in several versions: a first estimate, a revision, and figures recalculated when a country updates its census. A model trained on web pages has seen all of those versions, often copied into news stories and reports without dates. When it is asked for “the” number, it has no reliable way to know which one is current.
Definitions add a second layer of confusion. Child mortality, extreme poverty and school enrolment all have precise technical meanings, and small differences in wording produce different figures. A model that blends several sources can produce a number that matches none of them.
Hedging is a design choice
The high share of non-answers also reflects how chatbots are tuned. Developers train models to avoid stating figures they are unsure of, which is sensible for safety but frustrating for a user who wants a number. Our explainer on how AI models decide not to answer covers that trade-off. Connecting a model to a live, official source removes the need to guess, which is exactly the gap the UN is trying to close.
The caveats
The study is a UNICEF working paper being prepared for a journal and has not been peer-reviewed. UNICEF says it will publish the method, code and data alongside the paper. Until then, treat 21.2% as a strong signal rather than a settled figure.
How Data Commons Works
Data Commons is not a new database of UN numbers. It is a layer that connects existing datasets so they can be searched and combined.
A knowledge graph, not a spreadsheet
According to Google, the platform “automatically integrates metrics, timelines, and geographic boundaries into a single interconnected environment”. In practice, “Kenya”, “under-five mortality” and “2023” become linked entities, so a figure from UNICEF and a figure from the World Health Organization can sit side by side without a person reformatting either.
Plain-language search
Users can type questions such as how access to clean water in rural areas affects school attendance, or how life expectancy has changed across regions. The system returns relevant data and interactive charts. Google says this means a “nonprofit program manager”, a journalist or a policy analyst can use it without learning a database.
Provenance on every figure
The platform records where each statistic comes from, so a number retrieved by an AI system can be traced back to the original UN source. Azevedo said that matters more as people rely on AI tools to find and interpret information.
Why the silos mattered
Google says UN statistics “have lived in separate silos, organized in conflicting formats across, and within, different UN system organizations”, and that connecting them “often meant months of painstaking manual work for data analysts before any real analysis could begin”. Data Commons exists to remove that step.
How AI Agents Use Data Commons Through MCP
The feature that earned the headline is support for MCP, which lets AI systems connect to the data directly rather than guessing from memory.
What MCP is
The Model Context Protocol is an open standard for connecting AI applications to outside tools and data sources. Instead of each developer writing custom code for each database, an agent that speaks MCP can discover what a server offers and query it in a standard way. We explained a similar move in our piece on Google Home opening to any AI agent through MCP.
The Data Commons MCP server
Google publicly released the Data Commons MCP Server on 24 September 2025. It said the server gives “a standardized way for AI agents to consume Data Commons natively”, without developers learning the underlying APIs. It fits Google’s Agent Development Kit and Gemini CLI, and “can also be easily integrated with any other agentic workflow or platform”.
The first real use
The first public use case was the ONE Campaign’s ONE Data Agent, which lets users search “tens of millions of health financing data points in seconds, using plain language”. That tool shows the pattern the UN now wants to repeat at system scale.
The PEPFAR demonstration
At the launch, Google asked an AI system to find the impact of the US President’s Emergency Plan for AIDS Relief in Africa. Connected to the UN Data Commons through MCP, it found indicators on HIV infections, AIDS mortality and life expectancy and turned them into an infographic, without anyone locating or merging the datasets by hand.
UNdata Versus the UN System Data Commons
The change is easiest to see side by side.
| Feature | Old UNdata portal | UN System Data Commons |
|---|---|---|
| How you search | Browse and query a database | Plain-language questions and an Explore tab |
| Combining agencies | Manual download and merge | Linked in one knowledge graph |
| AI access | Scraping or custom code | MCP, a standard agent connection |
| Source tracking | Per dataset | Per statistic, traceable to the UN source |
| Outputs | Tables and files | Charts, dashboards and draft reports |
| Coverage goal | Selected statistics | 80% of UN system datasets by 2027 |
What stays the same
The underlying numbers are still produced by the same agencies using the same methods. Data Commons changes how they are reached, not how they are measured.
What changes for analysts
The biggest gain is time. Google says unified data gives analysts “more time to focus on uncovering key trends and designing evidence-based solutions, instead of formatting spreadsheets”.
AI Traffic Is Already Arriving
The UN is not preparing for a hypothetical. UNICEF’s own figures show AI assistants are already a major route to its data.
UNICEF’s data site
UNICEF’s data website gets more than 6 million visits a month and is among the agency’s most popular sites. Azevedo told TechCrunch that visits from users clicking links in ChatGPT answers rose 67% year on year between 1 January and 14 September 2026.
Bar widths are the percentages themselves. “About 10%” is UNICEF’s estimate of “about one in 10 visits”.
What those numbers imply
At 6 million visits a month, 6.4% is roughly 384,000 visits from ChatGPT links alone, by simple multiplication. If AI assistants are already sending that much traffic, making sure they quote the right number is a practical problem, not a theoretical one.
From SDG Portal to System-Wide Data Commons
The launch is the latest step in a partnership that has been building for three years.
| Date | Milestone |
|---|---|
| May 2018 | Google launches the Data Commons website |
| 2023 | Collaboration with the UN Statistics Division produces the UN Data Commons for the SDGs |
| September 2024 | Work with the UN International Computing Centre to scale across agencies including WHO, ILO and UNICEF |
| 24 September 2025 | Google releases the Data Commons MCP Server |
| 27 April 2026 | UN80 progress record lists the platform as Action 73, with 25 entities committed |
| 17 September 2026 | Public launch of the UN System Data Commons at data.un.org |
| 2027 | Target: 80% of UN system statistical datasets on the platform |
Why the SDGs drove it
The first UN version was built for the Sustainable Development Goals. When Google announced the expansion in 2024, it noted that the UN’s latest SDG report showed more than 80% of the goals were off track, with six years left to the 2030 deadline. Stefan Schweinfest, then director of the UN Statistics Division, said the platform “doesn’t just provide access to crucial information; it democratizes insights”. Tracking progress on poverty, hunger and clean energy needs data from many agencies at once, which is exactly what a linked system provides.
Part of UN80
The project sits inside the UN80 reform initiative as Work Package 16, led by the UN’s policy chief, the head of the Department of Economic and Social Affairs and UNICEF’s executive director. Action 73 calls for a shared platform so public data “can be accessed in one place (data.un.org) and used reliably with AI tools”.
A small discrepancy
The UN80 progress record from April says 25 entities had committed. TechCrunch reports 26 at launch. The difference is most likely one late addition, but it is a reminder to cite the date with the count.
What Could Go Wrong
A single platform for the world’s official statistics raises fair questions, and the people behind it acknowledge some of them.
Authoritative data, fallible conclusions
“Because models can misinterpret nuance, a human should always review the outputs before citing or publishing them,” Ramaswami told TechCrunch. Google’s own announcement says: “Even with grounded, verified data, review the underlying sources before citing critical figures.” Grounding cuts one kind of error, not all of them.
Dependence on one vendor’s platform
Critics may worry about the UN’s data running on Google technology. Two facts reduce that risk: Data Commons is open source, and the UN instance is UN-governed and designed to be run independently. The long-term test is whether UN staff can maintain it without Google engineers.
Definitions still differ
Linking datasets does not make their definitions identical. Two agencies can measure poverty or employment in slightly different ways. A knowledge graph can place them side by side, but a person still has to decide whether they are comparable.
Security and integrity
Official statistics shape aid budgets, investment decisions and elections, so the platform is a natural target. An open MCP endpoint that feeds many AI agents raises ordinary cybersecurity questions: who can change a dataset, how changes are logged, and how an agent can confirm it is talking to the real UN server. The UN has not published a detailed security model for the new platform yet, and it should.
Coverage is not complete
Data from nearly 20 entities is live, against a target of 80% of datasets by 2027. Until coverage is complete, an agent may find a figure from one agency and miss a better one from another.
How to Connect an Agent in Practice
For developers, the practical question is how to use this today. Google’s documentation for the Data Commons MCP Server points to a few routes.
Choose a client
The server works with Google’s Agent Development Kit and Gemini CLI, and Google says it can be integrated with any other agentic workflow. Because MCP is an open standard, most modern agent frameworks that support MCP can connect to it without custom code.
Ask narrow questions
Agents do best with specific requests: a named indicator, a named country and a year. “Under-five mortality in Kenya, latest year, with source” gives the agent far less room to misread than “how are children doing in East Africa”.
Always return the source and date
Instruct the agent to report the dataset, the agency and the reference year alongside every number. The platform keeps that provenance, so there is no reason to throw it away at the last step.
Spot-check against the website
For anything that will be published, open the figure on data.un.org and confirm it. That is the human review Ramaswami recommends, and it takes seconds once the agent has supplied the source.
What It Means for Businesses and Developers
The UN launch is a model for anyone who publishes or relies on authoritative numbers.
Ground agents in official sources
If your AI tools answer questions involving public statistics, connect them to a source like the UN Data Commons through MCP rather than trusting what the model remembers. The UNICEF result suggests memory alone is unreliable.
Keep provenance with every figure
Build your own reports so every number carries its source. That is what makes the UN system trustworthy, and it is what auditors and customers will ask for.
Test your own model
Run a small version of the UNICEF test on the questions your staff actually ask. Ask each one twice, days apart, and compare the answers.
Use it for research and reporting
Consultancies, NGOs, investors and sustainability teams already pull UN figures for market research, country risk reviews and ESG reports. A single search box across agencies, with charts and sources attached, can cut hours from that work. It also gives a defensible citation when a client or auditor asks where a number came from.
Watch your own AI referral traffic
UNICEF could only report its 67% rise because it measured referrals from AI assistants. Most analytics tools can separate that traffic today. Knowing how many visitors arrive from chatbots tells you how much your published figures matter to AI answers.
Make your own data AI-ready
Organisations that publish data should consider the same steps: structured formats, clear metadata and an MCP endpoint. As AI assistants send more traffic, being the source they quote becomes a competitive advantage. For more on this area, see our AI models, tools and releases hub.
Data Commons FAQ
What is the UN System Data Commons?
A platform at data.un.org, built on Google’s open-source Data Commons, that links statistics from across UN agencies and lets people and AI agents search them in plain language.
Is it free to use?
Yes. It is a public website and the underlying Data Commons software is open source.
What does MCP support mean?
AI agents that use the Model Context Protocol can connect to the platform and fetch official statistics and their sources directly, instead of relying on what the model learned in training.
How accurate were AI models without it?
In UNICEF’s unreviewed benchmark of six models and more than 133,000 responses, average accuracy was 21.2%, and about three in five answers gave no usable number.
When will all UN data be on it?
The target is 80% of the UN system’s statistical datasets by 2027. At launch, data from nearly 20 entities was available, out of 26 that have committed.
Does Google control the data?
Google.org funded the build with $2 million and Google engineers helped set it up, but the instance is UN-governed and meant to be run independently by the UN.
References
UN turns to Google to make its global data ready for AI agents (TechCrunch)
UN System Data Commons announcement (Google)
UN System Data Commons (data.un.org)
UN80 Initiative Work Package 16: UN System Data Commons, progress and next steps
Data Commons MCP Server release (Google for Developers)
UN Data Commons for the SDGs expansion (Google, 2024)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.