The latest

The United Nations and Google are creating a common platform allowing people and artificial intelligence systems to search, combine and use statistics across the UN. Built on Google’s open-source Data Commons technology, the UN System Data Commons allows users to bypass fragmented searches across traditional databases with natural-language questions and direct retrieval from official sources, while preserving each statistic’s origin. The initiative comes as generative AI becomes a primary gateway to information amid concerns over inaccurate or fabricated data.

Details

• Direct AI access: The platform supports the Model Context Protocol, or MCP, enabling AI systems to connect with external sources instead of reproducing statistics from training. Agents can query updated UN datasets and retrieve numbers with their original sources, allowing users to trace figures to responsible agencies.
• Low model accuracy: An ongoing UNICEF study found average accuracy of 21.2% among six prominent AI models answering more than 133,000 questions on global development indicators. UNICEF Chief Statistician João Pedro Azevedo said around three in five responses failed to produce a usable number. When models repeatedly returned one, they supplied the same figure only about half the time.
• Six systems tested: Tests 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. The research has not undergone peer review. UNICEF plans to publish its methodology, code and data with the final paper.
• Participation and funding: Twenty-six UN entities have committed to contributing data. Google.org provided $2 million and technical support for the platform’s infrastructure and development. Google said the system is hosted on an instance managed by the UN and designed to be operated and expanded independently by the organization.
• Expanded applications: Google demonstrated how an AI system connected through MCP could combine indicators, generate charts and dashboards, and produce written analysis without users manually merging datasets. Asked to assess the US President’s Emergency Plan for AIDS Relief in Africa, it retrieved UN data on HIV infections, AIDS-related mortality and life expectancy, then generated an infographic.
• Human review: The UN and Google stressed that authoritative statistics do not automatically make AI-generated conclusions authoritative. Models can misread context, correlations or nuances. Human verification remains essential before findings are cited, published or used to shape decisions.

Background

Google launched Data Commons in 2018 to organize public datasets from different sources within a shared framework. The UN partnership adapts it for large language systems to retrieve official real-world information dynamically instead of relying solely on training data.

What’s next

Datasets from nearly 20 UN entities are expected at launch. The UN targets integration of 80% of its statistical datasets by 2027, while UNICEF plans to release its final study with the methodology, code and underlying data.