联合国携手 Google，让全球数据为 AI Agent 做好准备

联合国周四宣布，正与 Google 合作，使其庞大的全球统计数据集合更便于 AI 系统访问和使用。

这一新系统名为 UN System Data Commons，基于 Google 开源的 Data Commons 平台构建，用户可通过自然语言查询检索来自各个联合国机构的统计数据。它将取代现有的 UNData portal；在该门户中，用户主要仍需通过较为传统的数据库界面浏览和搜索统计信息。新平台还支持模型上下文协议（Model Context Protocol，MCP），这是一项允许 AI 系统直接连接外部数据源的标准。

用户正越来越多地借助 AI 工具获取答案，但许多系统仍难以可靠地呈现权威数据。联合国儿童基金会（UNICEF）首席统计师 João Pedro Azevedo 在一场线上新闻简报会上告诉记者，UNICEF 对六款大语言模型进行了基准测试，涵盖超过 133,000 份针对全球发展指标问题的回答，平均准确率仅为 21.2%。

Azevedo 告诉 TechCrunch，该测试涵盖 OpenAI 的 GPT-4o 和 GPT-4o-mini、Anthropic 的 Claude Sonnet 4.5 和 Haiku 4.5，以及 Google 的 Gemini 2.5 Flash 和 Gemini 2.0 Flash。

Azevedo 表示，大约五分之三的回答根本未能提供可用数字，原因通常是模型对答案采取了模糊保留的表述。不过，当研究人员约两天后在相同模型版本上再次运行同样的问题时，那些两次都给出数字的模型，仅约半数情况下返回了完全相同的数字。

这项研究是一篇正准备投稿期刊的 UNICEF 工作论文，尚未经过同行评审。该组织表示，计划在论文发布时一并公开其方法论、代码和数据。

UNICEF 还发现，今年从生成式 AI 助手导向其数据网站的流量显著增加。该网站每月访问量超过 600 万次，是该机构最受欢迎的网站之一。Azevedo 告诉 TechCrunch，在 1 月 1 日至 9 月 14 日期间，用户通过 ChatGPT 回答中的链接访问该网站的次数同比增长 67%。这类引荐流量占今年全部会话量的 6.4%，而 UNICEF 估计，整体而言，AI 助手如今约占该网站每 10 次访问中的 1 次。

联合国表示，其 26 个实体已承诺加入 Data Commons，其中近 20 个实体的数据将在平台上线时提供。此外，联合国的目标是在 2027 年前将联合国系统中 80% 的统计数据集纳入该平台。

UN System Data Commons。图片来源： Google

联合国统计司代理司长 Shantanu Mukherjee 表示：“无论是在规模、范围还是灵活性上，我们都实现了数量级的进步，首次在整个联合国系统中连接起如此多的机构。我们也正借此机会让我们的数据为 AI 做好准备。”

Google.org 提供了 200 万美元的能力建设资金和技术支持，用于建立该平台的核心基础设施。领导 Google Data Commons 团队的 Prem Ramaswami 告诉 TechCrunch，该系统托管在由联合国治理的实例上，其目标是最终由联合国独立维护、运营并扩展。

Ramaswami 表示：“在整个推广过程中，我们采用了‘培训培训者’的方法，而且我们已经看到联合国系统团队迅速提升了能力。”

Google 于 2018 年推出 Data Commons，旨在将不同来源的公共数据集组织到一个通用框架中。去年，该平台新增了对 MCP 的支持，使 AI Agent 能够直接向 Data Commons 查询统计数据及其来源。

联合国的平台还会追踪每项统计数据的来源，因此人们可以将 AI 系统检索到的数据追溯至原始联合国来源。Azevedo 告诉记者，随着越来越多人依赖 AI 工具查找和解读信息，这一点十分重要。

除支持 AI Agent 检索单项统计数据外，Google 还演示了一个通过 MCP 接入联合国数据的 AI 系统，如何汇集多个指标，并在用户无需手动查找和整合底层数据集的情况下，生成仪表盘、图表和书面分析。

在一次演示中，Google 要求一个 AI 系统查找美国总统防治艾滋病紧急救援计划（PEPFAR）在非洲产生的影响。该系统识别出与 HIV 感染、艾滋病死亡率和预期寿命等指标相关的联合国统计数据，并利用这些数据制作了一张信息图。

不过，向 AI 系统提供权威数据，并不必然意味着其结论同样具有权威性。Ramaswami 表示：“由于模型可能误解细微差别，人类在引用或发布输出结果之前应始终进行审核。”

The United Nations on Thursday announced that it is working with Google to make its vast collection of global statistics easier for AI systems to access and use.

Called the UN System Data Commons, the new system is built on Google’s open source Data Commons platform and lets people search for statistics from across UN agencies using natural-language queries. It replaces the existing UNData portal, where users largely had to browse and search for statistics through a more traditional database interface. The new platform also supports the Model Context Protocol (MCP), a standard that allows AI systems to connect directly to external data sources.

Users increasingly turn to AI tools for answers, but many systems still struggle to reliably surface authoritative data. A UNICEF benchmark of six large language models across more than 133,000 responses to questions about global development indicators produced an average accuracy score of just 21.2%, João Pedro Azevedo, the agency’s chief statistician, told reporters in a virtual briefing.

The test 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, Azevedo told TechCrunch.

About three in five responses did not provide a usable number at all, often because the models hedged their answers, Azevedo said. However, when the same questions were run again on the same model versions about two days later, models that provided a number both times returned the identical number only about half the time.

The study is a UNICEF working paper being prepared for journal submission and has not yet been peer-reviewed. The organization said it plans to release its methodology, code, and data alongside the paper.

UNICEF has also seen a sharp rise this year in traffic from generative AI assistants to its data website, which receives more than 6 million visits a month and is among the agency’s most popular websites. Visits from users clicking links in ChatGPT answers to the site rose 67% year-over-year between January 1 and September 14, Azevedo told TechCrunch. Such referrals accounted for 6.4% of all sessions this year, while UNICEF estimates that AI assistants overall now account for about one in 10 visits.

The UN said 26 of its entities have committed to the Data Commons, with data from nearly 20 available at launch. Moreover, it aims to bring 80% of the UN system’s statistical datasets onto the platform by 2027.

UN System Data Commons.Image Credits: Google

“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.”

Google.org provided $2 million in capacity-building funding and technical support to establish the platform’s core infrastructure. Prem Ramaswami, who leads Google’s Data Commons team, told TechCrunch that the system is hosted on a UN-governed instance and is intended to eventually be maintained, operated, and scaled independently by the UN.

“We have taken a “train-the-trainer” approach throughout the rollout, and we have already seen the UN system team ramp up quickly,” Ramaswami said.

Google launched Data Commons in 2018 as an effort to organize public datasets from different sources into a common framework. Last year, it added support for MCP, allowing AI agents to directly query Data Commons for statistics and their sources.

The UN’s platform also keeps track of where each statistic comes from, so people can trace data retrieved by an AI system back to the original UN source. Azevedo told reporters that it was important as more people rely on AI tools to find and interpret information.

Alongside enabling AI agents to retrieve individual statistics, Google demonstrated how an AI system connected to the UN data through MCP could pull together multiple indicators and use them to generate dashboards, charts, and written analysis without a user having to manually find and combine the underlying datasets.

In one demonstration, Google asked an AI system to find the impact of the U.S. President’s Emergency Plan for AIDS Relief in Africa. The system identified relevant UN statistics on measures such as HIV infections, AIDS mortality, and life expectancy, and used them to produce an infographic.

However, giving an AI system authoritative data does not necessarily make its conclusions authoritative. “Because models can misinterpret nuance, a human should always review the outputs before citing or publishing them,” Ramaswami said.