🔬 applied physics

Polarization and Integration in Global AI Research

This study analyzes three decades of global AI research data to reveal a growing polarization between US and China-led innovation poles, where the UK and Germany align exclusively with the US while many other nations increasingly integrate with China, highlighting the complex interplay between geopolitical tensions and scientific collaboration.

Original authors: Luca Gallo, Riccardo Di Clemente, Balázs Lengyel

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Luca Gallo, Riccardo Di Clemente, Balázs Lengyel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of Artificial Intelligence (AI) research as a massive, bustling global marketplace. For decades, this marketplace was a place where scientists from every country would meet, swap ideas, build things together, and learn from one another. It was a giant, interconnected web of knowledge.

However, this paper tells the story of how that marketplace is slowly splitting into two distinct, rival neighborhoods, with a few clever traders trying to keep the doors open between them.

Here is the breakdown of what's happening, using simple analogies:

1. The Two Giants: The US and China

Think of the United States and China as the two biggest, most powerful store owners in this marketplace.

The Past: In the 1990s and early 2000s, they were neighbors who frequently visited each other's shops, shared recipes, and co-owned some projects.
The Present: Over the last 20 years, they have started building high fences. They are still the biggest players, but they are talking to each other much less than they should be, given how much they produce. They are drifting apart, creating two separate "poles" or centers of gravity.
- The US is tightening its rules, making it harder for Chinese researchers to visit or collaborate.
- China is building its own massive network, inviting countries that might feel left out by the US to join its side.

2. The Three Types of Neighbors

As these two giants pull apart, the rest of the world's countries have to decide where to stand. The researchers found that countries fall into three main groups:

The "US-Only" Club (The UK and Germany):
Imagine the UK and Germany as loyal friends of the US. They have tightened their ties with the American side and are actively stepping away from the Chinese side. They are choosing to hang out exclusively in the US neighborhood.
The "China-Only" Club (Developing Nations & Some Powers):
Many developing countries, and even some established ones like India and South Korea, are finding that the US door is closing. So, they are running toward China. They are building strong bridges with China and letting their connections with the US fade. It's like they are saying, "If the US won't play with us, we'll play with China."
The "Double-Booked" Bridges (France, Italy, Japan, Canada, etc.):
This is the most interesting group. These countries are like diplomats or bilingual translators. They are smart enough to keep a foot in both camps. They are collaborating with the US and China at the same time.
- Why does this matter? If the US and China stop talking entirely, the world's AI knowledge could get stuck in two separate bubbles. These "bridge" countries are the only ones keeping the flow of ideas moving between the two giants. They are the safety net preventing the whole system from breaking.

3. The European Puzzle

Europe is a bit of a mess in this story. While the European Union (EU) wants to be a united front, the paper finds that the big European AI hubs (like the UK, Germany, France, Italy) are actually drifting apart from each other.

Instead of a tight-knit European team, it looks more like a "hub-and-spoke" wheel. The big countries are connecting to their smaller neighbors, but they aren't connecting well with each other.
However, because some of these European countries are also the "Double-Booked" bridges, they have a unique chance to save the day. If they can stay neutral and keep talking to both the US and China, they could prevent the world from splitting into two hostile camps.

4. The Big Warning

The authors are worried. They say that if the US and China continue to drift apart and the "bridge" countries stop being able to talk to both sides, the global AI system will polarize.

The Risk: Imagine if the US only invented AI for the West, and China only invented AI for the East. We would end up with two different sets of rules, two different safety standards, and two different technologies that don't work together. This makes it very hard to regulate AI safely or ethically on a global scale.
The Hope: The "bridge" countries are the key. If they can maintain their connections to both sides, they can help keep the world integrated. But if political pressure forces them to pick a side, that bridge could collapse.

The Bottom Line

The world of AI is currently turning into a Cold War-style standoff between the US and China. While some countries are forced to pick a side, a lucky few are managing to stay in the middle. The future of safe, global AI depends on whether those "middle" countries can keep the lines of communication open before the two giants build walls that are too high to climb.

1. Problem Statement

The rapid advancement of Artificial Intelligence (AI) has become a focal point of geopolitical competition, particularly between the United States and China. While the US has adopted restrictive policies (e.g., the "China Initiative," mobility restrictions) to protect its technological advantage, China has pursued global governance leadership (e.g., WAICO).

The Core Question: How has the structure of international AI research evolved over the last three decades? Is the global system polarizing into two competing spheres of influence (US vs. China), or is there significant integration and bridging by third-party nations?
The Gap: Existing literature often focuses on bilateral counts or specific policy impacts. There is a lack of large-scale, network-based analysis quantifying the significance of cross-country relationships relative to random expectations, specifically distinguishing between mere volume and structural integration.

2. Methodology

The authors employed a network science approach using large-scale bibliometric data to analyze the evolution of the global AI research landscape from 1990 to 2023.

A. Data Collection

Source: OpenAlex (open-source bibliographic database).
Scope: ~1.99 million papers classified as "Artificial Intelligence" (1990–2021) and ~1.24 million non-AI papers citing them (up to 2023).
Granularity: Data was aggregated at the country level, tracking co-authorship (collaboration) and citation flows (knowledge exchange).

B. Network Construction

Two dynamic networks were constructed for each year:

Collaboration Network: Undirected, weighted. Nodes = Countries; Edge weight = Number of co-authored papers.
Knowledge Exchange Network: Directed, weighted. Nodes = Countries; Edge weight = Number of citations from country $i$ to country $j$ .

C. Statistical Significance via Null Models

To avoid biases caused by the sheer volume of publications (e.g., a large country naturally having more links), the authors compared observed link weights against randomized null models to determine statistical significance ( $z$ -scores).

Primary Model: Enhanced Configuration Model (ECM). A maximum-entropy approach that preserves the degree (number of partners) and strength (total volume of collaborations/citations) distributions of the empirical network while randomizing other structural features.
- Significance Metric: $z = \frac{w_{obs} - E[w_{null}]}{\sigma[w_{null}]}$ . Links with $z > 2$ (2 standard deviations above expectation) are considered "significant."
Robustness Checks: The authors validated findings using:
1. A variance-corrected maximum-entropy approach.
2. Local Rewiring Algorithm (LRA): Preserves node strengths in every individual realization (microcanonical framework) to ensure unbiased sampling.

D. Trajectory Analysis and Clustering

Temporal Evolution: The authors tracked the $z$ -score trajectories of country pairs from 2000 to 2021.
Clustering: Using k-means clustering with Dynamic Time Warping (DTW) as the distance metric, they identified four archetypal dynamics:
1. Scientific Convergence (SC): Increasing significance in both collaboration and knowledge exchange.
2. Scientific Divergence (SD): Decreasing significance in both.
3. Knowledge Integration (KI): Increasing knowledge exchange significance, decreasing collaboration.
4. Knowledge Divergence (KD): Decreasing knowledge exchange, stable collaboration.

3. Key Results

A. The US-China Divergence

Bilateral Trend: Despite the US and China having the highest absolute volumes of collaborations and citations, their relationship is statistically diverging.
Significance: The observed number of collaborations and citations between the US and China is significantly lower than what would be expected given their massive research outputs. This gap has widened since 2000.
Network Topology:
- US: Maintains significant ties with 46 countries across diverse continents (South America, Europe, Africa, Asia).
- China: Has significant ties with only 17 countries, mostly geographically proximate or historical allies (e.g., Pakistan, Vietnam, Hong Kong).
- Crucial Finding: The link between the US and China is not significant in the null model, indicating a decoupling of their research ecosystems.

B. Global Polarization Patterns

The study identified distinct integration pathways for other nations:

US-Aligned Divergence: The UK and Germany (major AI producers) are increasingly converging with the US while diverging from China.
China-Centric Convergence: A broad group of developed and developing nations (e.g., Brazil, Russia, Spain, Indonesia) is converging exclusively with China. Notably, countries like India, Poland, Singapore, South Korea, and Taiwan are simultaneously diverging from the US, signaling a shift in their research gravity toward China.
The "Bridge" Nations: A critical group (France, Italy, Netherlands, Canada, Japan, Israel) is doubly integrated, maintaining significant convergence with both the US and China. These nations act as structural bridges in the global network.

C. European Dynamics

Internal Fragmentation: The EU-28 lacks internal integration; major producers (UK, Germany, Italy, France) do not significantly collaborate with each other relative to their output.
Hub-and-Spoke: The European network is evolving into a hub-and-spoke structure where major producers connect to smaller neighbors but remain poorly integrated with one another.
External Alignment: European countries are split between the US pole, the China pole, and the "bridge" category.

4. Key Contributions

Methodological Innovation: The application of maximum-entropy null models to bibliometric data to distinguish between volume and structural significance. This reveals that high absolute numbers do not equate to strong integration if they fall below statistical expectations.
Quantification of Polarization: Provides empirical evidence that the US-China AI race is not just political but is structurally manifesting as a bipolar research network, with a statistically significant "cold war" in scientific collaboration.
Identification of "Bridge" States: Highlights the strategic importance of a specific group of countries (mostly European) that maintain dual integration. These nations are critical for preventing total fragmentation of the global AI ecosystem.
Dynamic Trajectory Mapping: Moves beyond static snapshots to show the velocity and direction of relationship shifts over 20+ years, identifying emerging poles (e.g., India's potential rise).

5. Significance and Implications

Policy & Regulation: The polarization threatens the effectiveness of global AI regulation. Fragmented regulatory frameworks (US vs. China) may hinder the development of ethical, broadly beneficial AI. The "bridge" nations are essential for interoperable standards.
Strategic Competition: For the US and China to "win" the AI race, they must not only focus on domestic capabilities but also on retaining the trust of the "bridge" countries. A purely bloc-based approach risks isolating the most productive science, which often arises in integrated networks.
Future Outlook: The study suggests that while the global AI system is polarizing, it is not yet binary. The future trajectory depends on whether the bridging capacity of countries like France, Japan, and Canada is leveraged to sustain cross-bloc knowledge flows or eroded by geopolitical pressure.

In conclusion, the paper argues that while the US and China are structurally decoupling, the global AI research system remains complex, with a critical subset of nations acting as the glue that prevents total fragmentation. The sustainability of global AI governance depends on preserving these bridges.