AI Gets World Statistics Wrong 4 Times Out of 5 — the UN Just Handed Its Data to AI Agents

Ask six of the world's best chatbots how many people live in poverty in Bangladesh, or how infant mortality in Ethiopia has changed over a decade, and roughly four times out of five you will get something other than the right answer. That is the blunt finding of a new UNICEF benchmark: across more than 133,000 responses, six leading large language models scored an average accuracy of just 21.2% on questions about global development indicators.

On Thursday, the United Nations announced its answer. It is working with Google to make its vast collection of global statistics directly accessible to AI systems — a new UN System Data Commons, built on Google's open-source Data Commons platform, that replaces the aging UNData portal and lets AI agents query authoritative UN statistics through the Model Context Protocol (MCP), the open standard for connecting AI systems to external data sources.

The United Nations Headquarters in New York, photographed in August 2015. The UN's new System Data Commons, built with Google, replaces the UNData portal and aims to make the organization's statistics directly usable by AI agents. Photo: qwesy qwesy, via Wikimedia Commons (CC BY).

21.2%: the number that started this

The UNICEF study is the kind of number that reframes a conversation. It tested 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 on questions about development indicators — the bread-and-butter statistics of global policy. Average accuracy: 21.2%.

The details are worse than the headline. About three in five responses did not provide a usable number at all, often because the models hedged their answers, UNICEF chief statistician João Pedro Azevedo told reporters in a virtual briefing. And when the same questions were run again on the same model versions about two days later, models that produced a number both times returned the identical number only about half the time — meaning even the "answers" weren't stable.

Two caveats matter here. First, the study is a UNICEF working paper being prepared for journal submission and has not yet been peer-reviewed — treat the 21.2% as a striking early result, not settled science. Second, the organization says it plans to release its methodology, code, and data alongside the paper, which will let outside researchers test whether the failure is as broad as it looks.

But the direction of travel is what the UN is betting on: users are already asking AI tools, not databases, for facts. UNICEF's own site draws more than 6 million visits a month, and visits referred from ChatGPT are up 67% year over year (Jan. 1–Sept. 14) — AI assistants now account for roughly one in ten visits, according to UNICEF. If the AI gives wrong answers about child mortality or poverty rates, the damage lands on the real world.

The UN's answer: hand the agents the database

The new platform is designed to end the era of models guessing at UN statistics from their training data. Users — and agents — can search statistics from across UN agencies with natural-language queries instead of navigating the UNData portal's traditional database interface. Since 2008, UNData had brought together more than 60 million data points from dozens of specialized agency databases, but its manual browsing tools, structured SQL-like searches, and CSV downloads were built for human researchers, not neural networks.

With MCP support, an AI agent can query the UN database directly instead of relying on pre-trained internal weights or scraping unstructured web pages. The system records the exact provenance of each statistic, giving users a verifiable route back to the originating UN agency — a direct attack on the hallucination problem, at least for this slice of knowledge.

"We are orders of magnitude more advanced in scale, scope, and flexibility, connecting for the first time across so many agencies across the UN system," said Shantanu Mukherjee, acting director of the UN Statistics Division. "And [we are] taking this moment to also make our data AI-ready."

This isn't the UN's first project with Google

It would be a mistake to frame Thursday's announcement as the beginning of the relationship. Google and the UN have been building this together for years. In October 2023, at the SDG Summit, UN DESA's Statistics Division and Google.org launched the UN Data Commons for the SDGs — a first iteration that integrated Sustainable Development Goals data into a public tool with AI-powered search. In 2025, Google announced the expansion of that platform to more agencies, including the World Health Organization, the International Labour Organization, and UNICEF, with work through the UN International Computing Centre to scale it system-wide.

This week's launch is the next phase: a system-wide rollout, backed by MCP, that absorbs and replaces the legacy UNData portal rather than sitting alongside it. Google introduced its Data Commons initiative in 2018 to organize public datasets — from the U.S. Census Bureau, the CDC, and the World Bank — into a unified knowledge graph, mapping differing schemas for geography, time periods, and demographics onto a shared standard. The UN project applies that machinery to the UN's own fragmented statistical holdings, under a UN-governed instance.

What changes for the people who actually use this

At launch, nearly 20 UN entities had statistics available on the platform, with 26 committed to the transition. The UN plans to move 80% of the statistical datasets across the entire UN system onto the platform by 2027. The datasets support monitoring of the 17 Sustainable Development Goals, established by the UN General Assembly in 2015 and measured through 231 unique indicators tied to 2030 targets.

For researchers and policymakers, the pitch is a single entry point with natural-language search and provenance you can audit. For developers, it's an MCP server that plugs authoritative statistics directly into agentic workflows — the kind of grounding layer that could keep an AI-generated country report from inventing its figures. Google.org provided $2 million in capacity-building funding plus hands-on technical support, and Prem Ramaswami, who leads Google's Data Commons division, said the project used a "train-the-trainer" approach so UN personnel can manage, operate, and expand the infrastructure independently. The UN Statistics Division, which has standardized international statistical systems since 1947, oversees the shift.

It is also worth noting the timing: the announcement lands days before the UN General Assembly's high-level week, when world leaders — including a U.S. president who has spent the week clashing with the UN over AI governance — will be in the building. (The U.S.–China AI-risk talks expected on the summit's sidelines are here.) Data infrastructure rarely gets a starring role at UNGA, but this year the UN is betting that AI-ready statistics are diplomacy-adjacent.

The ground-truth arms race

Analysis — the following is interpretation, not reported fact. The UNICEF benchmark is the real story here, and the Data Commons is the UN's attempt to change what it measures. A 21.2% accuracy rate on development statistics is not a curiosity — it is a quiet crisis for anyone who treats AI assistants as oracles. Students writing papers, journalists checking figures, voters asking about the economy: all of them are downstream of models that, per this working paper, hedge or hallucinate most of the time on exactly the questions where being wrong has consequences.

The UN's move is best understood as a bid to become the ground-truth layer for the agentic era. If AI agents are going to answer factual questions at scale, someone has to supply the facts — and the institution with 60 million curated data points, a Statistics Division older than most of the world's statistical agencies, and no commercial incentive to fudge the numbers has a credible claim to the job. Google's role is the distribution: the MCP server, the Gemini CLI and Agent Development Kit integrations, the developer on-ramps that turn a UN database into something an agent can actually call.

The open questions are the ones no launch announcement can answer. Will model builders actually route factual queries through the Commons, or will convenience keep them guessing from weights? Will the 80%-by-2027 migration hit its target, given how slowly UN system reforms usually move? And will provenance survive the journey from API response to chatbot answer — or will the AI still summarize the verified number into something unverified? The UN has built the library. Whether the AI reads from it is up to the industry it is trying to fix.

One final note on sourcing: this story rests primarily on TechCrunch's Sept. 17 reporting, supplemented by the UN's and Google's published accounts of the earlier phases of the collaboration. No separate UN DESA or Google announcement of this week's system-wide launch was available at the time of writing; where this article describes the new platform, it is describing what was reported Thursday.

Sources: TechCrunch, "UN turns to Google to make its global data ready for AI agents" (Sept. 17, 2026); United Nations, "Empowering SDG actions with new tools for tracking SDG progress" (Oct. 2023); Google, "UN Data Commons for the SDGs expands to more agencies" (2025); Google, "New tool from Google and the UN to track progress on Global Goals" (2023).