UN launches AI-friendly data platform with Google to improve statistical accuracy
The United Nations unveiled the UN System Data Commons, built on Google's open-source platform, allowing natural-language queries across agency statistics and supporting the Model Context Protocol for AI connections. A UNICEF benchmark of six large language models found an average accuracy of just 21.2% on global development indicators, with many responses lacking usable numbers. The UN aims to bring 80% of its statistical datasets onto the platform by 2027.
The platform is hosted on a UN-controlled instance, with Google.org contributing $2 million for capacity building and technical support. Google's Data Commons, which began in 2018, added MCP support last year, enabling AI agents to query statistics directly. The UN intends to eventually manage the system independently, using a "train-the-trainer" approach.
The UNICEF study, a working paper not yet peer-reviewed, tested six models across over 133,000 responses. It found that 60% of responses lacked a usable figure, and even when numbers were provided, consistency across repeated runs was only about 50%. Meanwhile, referrals from ChatGPT to UNICEF's data site surged 67% year-over-year, with AI assistants now accounting for roughly one in ten visits.
This initiative could significantly improve the reliability of AI-generated answers about global development, potentially benefiting policymakers, journalists, and researchers who depend on accurate statistics. However, the low baseline accuracy of current models suggests that without such structured data access, AI tools may continue to propagate misinformation. The platform's traceability features could foster greater trust in AI-sourced figures, though it may also encourage over-reliance on AI assistants, potentially reducing direct human engagement with primary UN data sources.