AI’s Shadow Global Governance

Originally published on Project Syndicate on May 15, 2026.

Judging by most of the media coverage of the global AI "race," the United States and China are the two dominant players, and every other country and organization is at risk of being left behind. The reality is far more complex—and not nearly as bleak.

WASHINGTON, DC—In The Three-Arched Bridge, the novelist Ismail Kadare tells the story of a bridge being built in medieval Arbëria (modern-day Albania, Kadare’s homeland) just as the Byzantine Empire is giving way to the Ottomans. The bridge is exceedingly difficult to build and equally difficult to cross. With every week bringing new headlines about the fierce rivalry between American and Chinese AI developers, Kadare’s bridge is an apt metaphor for today’s global AI governance.

In the United States, policymakers are obsessed with weaponizing their country’s AI advantages. Congress and the White House both aspire to leverage America’s “compute power”—advanced semiconductors and data centers—while holding the rest of the world at a negotiating disadvantage. The narrative in both the US and China is that we live in a dog-eat-dog world where no bridges can be built.

As someone who leads a large team of interdisciplinary researchers using computational methods to analyze global AI governance, I believe the issue is more complex than the “great power rivalry” narrative suggests. Our approach traces the many interconnections among national and multilateral AI strategies, revealing commonalities, distinctions, and how states and organizations are learning from and influencing each other.

Such influence need not always flow from the strong to the weak. AI systems used by pastoralists in Africa would almost certainly prove relevant elsewhere. That is already true of India Stack, India’s identity and payments portal, which is being widely emulated across the developing world. As technologies and power diffuse globally, the weak are learning from each other and finding ways not to end up at a disadvantage.

Since 2016, more than 70 countries have published national AI strategies, as have the European Union and multilateral organizations such as the United Nations. Together, these documents offer a wealth of textual data for analysis, and my team has painted a granular picture of the topics these documents contain. Our topic modeling presents the distributions of words in the documents alongside their probability distributions, much like a large language model. Among our findings, three notable results stand out.

First, several countries cluster together around particular topics or priorities. For example, EU countries cluster around an approach that seeks to balance economic competitiveness with ethical and social concerns, and Latin American countries cluster around one that builds on existing digital infrastructure. By contrast, China and the United States do not cluster with any other states. Each has a unique strategy that is primarily concerned with global dominance, research and development, and scientific expertise.

Second, beyond regional clusters, countries also share their approaches through international networks. Thus, Spain shows up as a player in both the Latin American and EU clusters. Sweden clusters with the EU in one topic but also with Finland in another (namely, one that pairs enhanced social, labor, and welfare issues with a strategy favoring startups and economic reforms).

Figure 1: Topic models relating UN organization/country–topic distributions. The light-to-dark scheme visually conveys low-to-high probabilities of topics in a country or multilateral agency.

Adapted from: J P Singh, Manpriya Dua, Amarda Shehu. 2026. Diffusion of power and multiplexed governance: evolving networks and clusters for global governance of AI Infrastructures. International Affairs.  Volume 102, Issue 2, pp. 409-434.

Equally notable, multilateral organizations do not seem to cluster with countries. The World Health Organization, for example, produces documents only in the health topic, so the appropriate unit of comparison would be national AI health strategies. The exceptions are the EU cluster and the World Bank, whose approach toward AI shares much with many developing countries.

Of course, technologies always embody innovation and learning, which in turn inform subsequent rules and institutions. We should not be surprised that learning is taking place globally as AI diffuses and evolves. The surprise is in the headlines that focus solely on great-power competition and the “left behinds.” While those are valid concerns, they represent only part of the picture. They do not convey the complex learning mechanisms that make the World Bank show up centrally in developing-country clusters, or that show Spain’s AI thinking has much in common with how Latin American governments see the issue.

Here in Washington, it’s almost anathema nowadays to speak of “global cooperation” on AI governance. But there is nothing fanciful about the empirical reality of global influence and emulation. That reality is evolving as quickly as AI infrastructure itself, suggesting that something like global cooperation is already taking place.

Like Kadare’s bridge between civilizations, formal connections are still difficult to create and sustain, especially if they involve the two big AI powers. But the task is coming much easier to the rest of the world. Other countries and organizations have an opportunity to share what they know, to learn from each other, and ultimately to create shared institutions and standards that they can all live with.

J.P. Singh, University Professor at the Schar School of Policy and Government and Director of the Center for AI Innovation and Economic Competitiveness at George Mason University, is Co-Editor-in-Chief of Global Perspectives.