AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Concepts & Frameworks

AI Diffusion

medium confidence · updated 2026-08-28

The spread and adoption of AI systems globally — a key axis of competition between the US and China, determining whose values shape the digital order.

Contents

AI diffusion is the spread and adoption of AI systems across countries, industries, and institutions. In the context of US-China competition, diffusion is framed as the process that determines whose systems become the global default, and therefore whose values, governance norms, and data practices define the digital order. Geopolitics in the Age of Artificial Intelligence argues that diffusion may matter as much as or more than frontier capability, particularly in scenarios where breakthroughs are easy to copy.

Geopolitical dynamics

The argument advanced in Geopolitics in the Age of Artificial Intelligence is that the systems that take root first shape technical standards, create dependency, and define governance norms, and that this gives diffusion strategic weight independent of who holds the capability lead. In this account, early deployment creates lock-in through familiarity, integration, and switching costs, conferring a first-mover advantage. Beijing is described as treating AI governance itself as a strategic export, bundling systems with financing and development projects.

Several factors are identified as the substrate of diffusion. Infrastructure — data centers, networks, and regionally tailored systems — determines where systems can be deployed. Development finance shapes which markets receive systems at all: institutions such as the US Development Finance Corporation could fund deployment in markets the private sector will not serve. The choice between open-source and proprietary distribution is also treated as a diffusion lever; open-sourcing safe models is presented as potentially the best diffusion strategy when copying is easy, corresponding to Worlds 7 and 8 in the Eight Worlds Framework.

Domestic and organizational diffusion

Beyond cross-border diffusion, the spread of AI within organizations is treated as a separate problem. Stanford HAI research (Karunakaran et al.) found that corporate AI project success depends on three organizational variables rather than technical capability alone:

  • Jurisdictional clarity — whether there is a well-defined group of domain experts.
  • Task centrality — whether the target task is core to the experts' daily work.
  • Task enactment homogeneity — whether experts perform the task the same way.

According to the study, AI adoption succeeds when all three conditions are favorable and projects fail when they are not, even with the same development team. The finding is framed as evidence that AI readiness is as much a matter of organizational structure as of technology or talent supply. (Source: hai.stanford.edu)

Measurement

Diffusion is difficult to measure because there is no agreed definition of adoption. Cross-country comparisons turn on whether a study counts any use of a chatbot, use of AI in producing goods or services, or integration into production processes, and survey wording changes have moved reported figures on their own. A Brookings account of the evidence notes that U.S. Census data from February 2026 put roughly 17.5% of U.S. businesses as having used AI in at least one business function in the previous two weeks, and that the question was rewritten in November 2025 from earlier wording about using AI "in producing goods or services." Its stated conclusion is that only the direction is safe to infer: adoption is rising, uneven across countries and regions, and faster among larger firms and digitally mature sectors (Source: brookings.edu).

The most granular cross-country series is Microsoft's quarterly Global AI Diffusion Report, produced by its AI Economy Institute, which defines diffusion as the share of people aged 15 to 64 who have used a generative AI product, derived from aggregated and anonymized Microsoft telemetry and adjusted for operating-system and device market share, internet penetration and country population. The Q1 2026 edition put global usage at 17.8% of the working-age population, with 26 economies above 30%. The United Arab Emirates led at 70.1%, followed by Singapore at 63.4%, Norway at 48.6%, Ireland at 48.4% and France at 47.8%; the United States ranked 21st at 31.3%, up from 24th the previous quarter. Microsoft's chief data scientist Juan Lavista Ferres recorded "continued widening of the AI gap between the Global North and South, with usage now at 27.5% in the North and 15.4 percent in the South," a 12.1-percentage-point difference, with the South gaining 1.3 points from 14.1% over the quarter — roughly half the North's rate of increase. The company states that no single metric fully captures adoption (Source: microsoft.com; campustechnology.com).

The underlying metric, published as "Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage," combines the share of users engaging with AI, desktop device penetration estimates, and a country-specific mobile-to-desktop usage ratio. Its authors note that it spans 147 countries and regions, against Stanford HAI's Global AI Vibrancy Tool, which ranks 36 predominantly high- and upper-middle-income economies. The paper's stated finding is that usage among internet-connected populations in lower-income countries is substantially higher than country-level figures suggest, which the authors read as indicating that the binding constraint is internet access rather than demand; it also reports that the increase in usage flattens above a certain income level (Source: arxiv.org).

Barriers in lower-income countries

A World Bank analysis published in May 2026 sets out the material constraints behind the measured gap. Focusing on the most-used platform, it reports that visits to ChatGPT per internet user in high-income countries were nearly 50 times those in low-income countries as of May 2025, 8 times those in lower-middle-income countries and 3 times those in upper-middle-income countries; it notes this partly reflects different platforms being used in different markets, with OpenAI less present in China. Anthropic's own economic index is cited as showing country usage scores in poorer countries roughly 10 times smaller than in high-income countries. Measures accounting for multiple platforms find that in most Sub-Saharan African countries fewer than 10% of the working-age population uses AI.

The analysis identifies four constraints. Connectivity and cost: only a minority of low-income-country households can afford data, where a 5GB mobile package costs more than 15% of the average person's total budget and, with about 60% of household budgets going to food, would take up nearly half of what remains. Skills: fewer than 5% of the population of low-income countries were estimated in 2023 to have basic digital skills such as email, web search and online transactions. Language: English accounts for nearly half of global URLs and a disproportionate share of the high-quality training text, with a significant data deficit for languages spoken primarily in low- and middle-income countries. Compute: as of June 2025, high-income countries held 77% of global data-center capacity, middle-income countries 23% and low-income countries less than 0.1%, with the United States holding some 20,000 times as many servers per capita as low-income countries and 200 times as many as middle-income countries. The analysis notes that computing power is tradable across borders, so access can substitute for domestic capacity (Source: data360.worldbank.org).

On exposure at work, the same source reports that fewer than one in five workers in low-income countries is highly exposed to AI, that the gap is likely understated because occupations with the same title involve less AI-exposed tasks in poorer countries, and that exposure is concentrated in urban populations. In middle-income countries, ICT workers and teachers account for nearly three-quarters of all AI usage, against a broader spread in high-income countries (Source: data360.worldbank.org).

Relation to other concepts

Diffusion interacts with several other policy and technical concepts. It stands in tension with Export Controls (AI): controls can slow the diffusion of US systems abroad, potentially ceding ground to China. It depends on Compute Governance, because infrastructure determines where systems can be deployed. Its importance varies across the Eight Worlds Framework, where it is central in Worlds 5, 7, and 8 and important but secondary in Worlds 1 and 2. Emergent Misalignment complicates open diffusion strategies: if fine-tuning open models can produce dangerous behavior, then open distribution carries risk.

Relationships

Sources

Categories: Concepts & Frameworks | us | china | global | competition | framework | workforce | compute
What links here (36)