UN launches AI-ready Data Commons with Google as UNICEF benchmark finds six major LLMs score just 21.2% on global statistics

The UN and Google launched the UN System Data Commons, an open-source, MCP-enabled platform replacing the UNData portal so AI agents can retrieve traceable UN statistics; a UNICEF benchmark of 133,000+ responses found GPT-4o, GPT-4o-mini, Claude Sonnet 4.5,…

The UN has launched the UN System Data Commons, an open-source platform built on Google's Data Commons that replaces the UNData portal and unifies siloed UN statistics into a single AI-ready knowledge graph. It offers natural-language search with interactive visualizations and supports the Model Context Protocol (MCP), letting AI agents autonomously fetch authoritative figures traceable to original UN sources and assemble charts, infographics, or draft reports. Datasets are validated with UN statisticians, Google.org contributed $2 million in funding, and the UN-governed instance will run independently. The UN aims to include 80% of UN system statistical datasets by 2027; 26 UN entities have committed and data from nearly 20 is already available. The launch comes alongside a UNICEF benchmark of over 133,000 responses showing major LLMs struggle with global development data: GPT-4o, GPT-4o-mini, Claude Sonnet 4.5, Claude Haiku 4.5, Gemini 2.5 Flash, and Gemini 2.0 Flash averaged just 21.2% accuracy, roughly three in five answers provided no usable number, and repeat runs of identical questions returned matching numbers only about half the time. Demand signals reinforce the gap: ChatGPT referrals to UNICEF's data site rose 67% year over year.

• UN System Data Commons launched, built on Google's open-source Data Commons, replacing the UNData portal
• Platform unifies siloed UN statistics into one interconnected knowledge graph with natural-language search and interactive visualizations
• Model Context Protocol (MCP) support lets AI agents autonomously fetch UN figures traceable to original sources
• UNICEF benchmark of 133,000+ responses found six frontier LLMs averaged 21.2% accuracy on global development indicators
• Models tested: GPT-4o, GPT-4o-mini, Claude Sonnet 4.5, Claude Haiku 4.5, Gemini 2.5 Flash, and Gemini 2.0 Flash
• Roughly three in five benchmark answers provided no usable number; repeat runs matched only about half the time
• Google.org contributed $2 million; the UN-governed instance will run independently
• 26 UN entities committed; data from nearly 20 available; 80% of UN system statistical datasets targeted by 2027