UN and Google launch Data Commons for AI-ready statistics

The United Nations has announced the UN System Data Commons, a data platform developed with Google to make global statistics easier for people and AI systems to find and use. Built on Google’s open-source Data Commons platform, it replaces the UNData portal’s more traditional database browsing and search interface with natural-language queries.

Twenty-six UN entities have committed to the initiative, and data from nearly 20 is available at launch. The UN aims to place 80% of the UN system’s statistical datasets on the platform by 2027. Google.org provided $2 million in capacity-building funding and technical support for the core infrastructure.

A common layer for UN statistics

Google launched Data Commons in 2018 to organize public datasets from different providers into a common framework. The UN instance is governed by the UN and is intended eventually to be maintained, operated and scaled independently by the organization. Google said it has used a train-the-trainer approach during the rollout.

The platform supports the Model Context Protocol, or MCP, which enables AI systems to connect directly to external data sources. Data Commons added MCP support last year. On the UN platform, a retrieved statistic retains information about its origin, allowing users to trace an AI-accessed figure back to the relevant UN source.

Google demonstrated an MCP-connected AI system retrieving several UN indicators, then using them to create dashboards, charts and written analysis without the user manually locating and combining the underlying datasets. In one example, the system assessed the impact of the U.S. President’s Emergency Plan for AIDS Relief in Africa using measures including HIV infections, AIDS mortality and life expectancy.

Reliable retrieval remains an AI challenge

UNICEF tested six large language models on more than 133,000 responses to questions about global development indicators. The models included 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.

João Pedro Azevedo, UNICEF’s chief statistician, said the benchmark produced an average accuracy score of 21.2%. About three in five responses did not provide a usable number, often because the models hedged. When identical questions were run again on the same model versions roughly two days later, models that supplied a number both times returned the identical figure only about half the time. The UNICEF working paper is being prepared for journal submission and has not yet been peer-reviewed; UNICEF plans to release its methodology, code and data with the paper.

Demand for AI-mediated access is also growing. UNICEF’s data site receives more than six million monthly visits, and visits from people following ChatGPT links rose 67% year over year between January 1 and September 14. Those referrals represented 6.4% of sessions, while UNICEF estimates AI assistants account for about one in 10 visits.

Traceability does not remove the need for review

Providing models with authoritative source material does not make every interpretation authoritative. Google’s Data Commons lead Prem Ramaswamy said human review remains necessary before AI-generated output is cited or published because models can misinterpret nuance. For businesses using AI to assemble evidence, the practical implication is to combine direct, traceable data access with a defined human review step for every statistic, visualisation and conclusion.