UN and Google launch AI-ready data commons to power global statistics access

The United Nations has teamed up with Google to rebuild how its global statistics are searched, shared and used by AI systems, launching a new platform designed to make UN data easier for both people and machines to access. The UN System Data Commons replaces the older UNData portal and is meant to help AI agents retrieve authoritative figures directly from UN sources, a change that matters as more users rely on chatbots for facts they increasingly get wrong.

The rollout comes as the UN system confronts a simple but urgent problem: artificial intelligence tools are becoming a default way to look up information, yet they still struggle to deliver reliable numbers on development, health and other public-interest topics. By tying its datasets to Google’s Data Commons and the Model Context Protocol, the UN is trying to ensure that answers generated by AI tools can be grounded in official statistics rather than guesswork.

Why the United Nations is rebuilding access to its data

The short answer is that the UN wants its data to be easier for humans to find and easier for AI systems to use without losing traceability. The new system is intended to make a sprawling collection of statistics from across UN agencies searchable through natural language rather than through a conventional database-style portal.

For years, users approaching UNData had to know how to navigate a structured repository and often had to piece together results from multiple agency sources. The new UN System Data Commons is designed to work more like a modern search layer: a user can type a question in plain English and receive relevant statistics, with the source trail preserved.

The platform also reflects a broader shift in information habits. People are increasingly asking chatbots for quick answers, including questions about poverty rates, maternal mortality, education access and disease prevalence. The UN sees that behavior as both a challenge and an opportunity: if AI tools are going to answer those questions, the underlying data should come from trusted, traceable sources.

What changed from UNData to the Data Commons?

The new system is not just a cosmetic refresh. It is a structural redesign that aims to connect data across many UN entities through a shared framework based on Google’s open-source Data Commons technology. That lets statistics from different agencies sit within a consistent model, making them easier to query, compare and reuse.

According to the UN, the platform is already connected to 26 entities, with data from nearly 20 available at launch. The long-term goal is more ambitious: the organization wants 80% of the UN system’s statistical datasets onboarded by 2027.

That scale is important because the UN is not a single data producer. It is a network of agencies, funds and programs, each with its own statistical holdings and publishing conventions. The new architecture is supposed to reduce fragmentation and make it possible to answer cross-cutting questions from a single entry point.

How does the system work with AI agents?

The system supports the Model Context Protocol, or MCP, a standard that allows AI applications to connect directly to external data sources. In practical terms, that means an AI agent can query the UN’s data repository without relying solely on what it learned during training or on an unverified web search.

MCP matters because it gives AI tools a way to fetch statistics, inspect sources and assemble outputs on demand. Instead of simply predicting a likely answer, a model can be instructed to retrieve the relevant numbers from an authoritative dataset and cite where they came from.

Google showed how that could work in a demonstration using UN data. In the example, an AI system was connected through MCP and asked to assess the effect of the U.S. President’s Emergency Plan for AIDS Relief in Africa. The model identified relevant indicators such as HIV infections, AIDS mortality and life expectancy, then used those figures to generate an infographic.

The UN and Google say the key is not only making data accessible, but making it traceable so people can see where each figure came from and check it against the original source.

The platform keeps provenance information attached to each statistic, which means a user or an AI-generated answer can be traced back to the underlying UN dataset. That is a critical feature at a time when AI-generated responses often present numerical claims with a confidence that the underlying models do not deserve.

Why accuracy remains a problem for AI-generated statistics

One of the clearest signals behind the UN’s move is that large language models still struggle badly when asked to produce reliable development data. UNICEF recently tested six major models across more than 133,000 responses to questions about global development indicators and found an average accuracy score of just 21.2%.

That means most responses were not dependable enough to be treated as factual answers. The test covered 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, according to UNICEF chief statistician João Pedro Azevedo.

Azevedo said about three in five responses failed to provide a usable number at all, often because the models hedged instead of committing to a value. In repeat tests conducted roughly two days later, some models that returned a number both times gave the same number only around half the time.

That inconsistency is not a minor technical flaw. For public statistics, reproducibility and precision are the entire point. If the same question yields different numbers depending on when or how it is asked, the answer cannot safely support policy, reporting or research.

What does the UNICEF benchmark tell us?

The UNICEF working paper suggests that the problem is not limited to isolated factual mistakes. The models often struggled to provide a direct number at all, and when they did, the answers were frequently unstable. That points to a broader weakness in how general-purpose AI systems handle structured statistical questions.

The study has not yet been peer-reviewed, but UNICEF said it plans to publish the methodology, code and data alongside the paper. Even before formal publication, the findings help explain why the UN is prioritizing AI-ready official data rather than assuming models can reliably “know” public statistics on their own.

What Google contributed to the UN project

Google’s role goes beyond being a vendor. The company’s open-source Data Commons platform provides the technical foundation, while Google.org supplied $2 million in capacity-building funding along with support for the system’s core infrastructure.

Google’s Data Commons was originally launched in 2018 to unify public datasets from different sources into a common framework. Last year, it added MCP support, which made it possible for AI agents to query statistics and retrieve source information directly from the platform.

That evolution lines up with the broader direction of enterprise AI: systems are moving away from purely generative chat toward tool-using agents that can call external databases, APIs and documents. In that model, a system’s value is no longer just in its language ability but in its ability to combine reasoning with trusted retrieval.

Prem Ramaswami, who leads Google’s Data Commons team, said the UN instance is hosted under UN governance and is intended to be operated, maintained and expanded by the UN over time. He described the rollout as a train-the-trainer effort, with the UN team already taking on more responsibility as the system matures.

Google says the long-term plan is for the UN to run the platform independently, after building internal capacity through the initial rollout.

How AI assistants are already changing traffic to UN data

AI tools are not just a future use case for the UN’s statistics; they are already reshaping traffic patterns. UNICEF said referrals from links surfaced in ChatGPT answers to its data website rose 67% year over year between January 1 and September 14.

Those referrals accounted for 6.4% of all site sessions during the year to date. UNICEF estimates that AI assistants overall now drive about one in 10 visits to its data site, which receives more than 6 million visits a month and ranks among the agency’s most popular properties.

That level of traffic matters for two reasons. First, it shows that AI systems are already becoming a major gateway to official information. Second, it demonstrates that the quality of those gateways matters: if AI assistants are going to funnel users to UN or UNICEF data, those users need to land on information that is accurate, contextualized and verifiable.

Why referrals from chatbots matter to public agencies

For public institutions, chatbot referrals can increase reach, but they also create new risks. A user may never visit the underlying source if the AI answer appears complete enough on its own. That makes provenance, citations and context more important than ever.

The UN platform’s ability to preserve links back to original statistics is meant to address that problem. If a chatbot summarizes a figure incorrectly, users should be able to inspect the source rather than taking the summary at face value.

How the UN plans to scale the system

The UN’s strategy is to gradually bring more agencies and datasets into the commons, using a shared framework that can support direct machine access. The organization said 26 entities have already agreed to participate, and data from nearly 20 are available now.

The target for 2027 is significant: the UN wants 80% of its statistical datasets to be available through the platform. Achieving that would require not only technical integration, but also organizational alignment across a highly distributed system.

This is where Google’s “train-the-trainer” approach becomes important. Rather than relying on outside engineers indefinitely, the project is intended to help the UN build internal expertise so it can expand the platform itself.

That approach also suggests a deeper ambition than a one-off product launch. The goal is to create durable institutional capacity that can keep pace with changes in AI tools, data standards and user expectations.

Why traceability is central to trustworthy AI data

The UN and Google are both emphasizing a point that often gets lost in conversations about AI: access is not the same as truth. Giving an AI system a better dataset improves the odds of a correct answer, but it does not guarantee one.

Ramaswami warned that models can still misunderstand nuance, which means a human should review outputs before they are cited or published. That caution is especially important when the output is meant to support policy discussions, journalism or research.

In other words, the new platform is best understood as an infrastructure layer, not an oracle. It can help AI systems fetch authoritative numbers and reduce hallucination risk, but it cannot eliminate the need for editorial judgment.

This distinction is likely to become increasingly important as more government agencies, multilaterals and research institutions explore agent-ready data systems. The core question is not whether AI can summarize statistics, but whether those statistics are structured well enough for machines to use without introducing new errors.

What this means for the broader AI ecosystem

The UN’s move is part of a wider trend toward “retrieval-first” AI, in which models are paired with trusted databases and tools. For companies, that trend is often about productivity and workflow automation. For international organizations, it is about credibility.

There is also a competitive dimension. The major AI platforms are racing to become the default interface for information. If official institutions do not make their data machine-readable and queryable, AI systems will still answer questions about that data, just without the institutions’ direct involvement.

By establishing its own AI-ready data layer, the UN is trying to retain control over how its statistics are discovered and interpreted. It is also creating a precedent for how public-sector information can be made safer for use in agentic systems.

• The UN System Data Commons replaces the older UNData portal with a search experience built for natural language and AI access.
• The platform uses Google’s open-source Data Commons and supports MCP so AI agents can query official statistics directly.
• UNICEF testing found major large language models performed poorly on development data accuracy, reinforcing the need for authoritative sources.
• Google.org contributed $2 million and technical support, while the UN aims to independently scale the system by 2027.
• AI assistants are already sending meaningful traffic to UN and UNICEF data sites, showing the shift is already underway.

Timeline of the rollout

Here is a concise look at how the UN and Google project has developed so far.

The bottom line

The UN’s partnership with Google is an attempt to solve a problem that is becoming central to the AI era: how to make trusted public data usable by machines without losing accuracy, attribution or oversight. The new System Data Commons is meant to do exactly that, turning a fragmented collection of statistics into a machine-readable resource for people and AI agents alike.

Whether it succeeds will depend on execution, adoption and the ability of the UN to scale the system across its many agencies. But the underlying message is clear: as chatbots become a primary interface for information, institutions that produce the world’s most important data can no longer afford to leave that data in formats machines struggle to use.