The AI Matrix Framework
Most measures of AI adoption count who has the tools. The AI Matrix asks a harder question: who gains agency from them? It separates access (how broadly AI reaches a population) from agency (how productively people engage with it). The difference matters because broad access without agency produces passive dependency, not empowerment. The framework comes from Simpson (2026) and is recomputed here each time Anthropic publishes new data.
Two dimensions
Access: how many people use AI
Access is the log-transformed per-capita rate of conversations with Claude. A high-access country has AI reaching a broad share of its population. A low-access country has usage concentrated among a few.
Agency: what they do with it
Agency is a composite of four indicators: how often people co-create with AI (task iteration, learning, validation) rather than just issuing directives, how often tasks succeed, and the education level reflected in prompts. High agency means people treat AI as a collaborator. Low agency means they treat it as a vending machine.
Four quadrants
Elite Empowerment
Low access, high agency. Few people have AI, but those who do use it productively. Argentina, Mexico, and Turkiye sit here.
Full Empowerment
High access, high agency. AI reaches a broad share of the population and people co-create with it. The target condition: not merely adoption, but genuine capability gain. The US, UK, Germany, Japan, and most of Western Europe sit here.
Full Dependency
Low access, low agency. AI barely reaches the population, and what use exists skews toward coursework or one-off directives. The capability overhang is widest here: the gap between what AI could do and what it actually does. India, Brazil, Kenya, Nigeria, and much of the Global South sit here.
Passive Dependency
High access, low agency. The tools are there, but usage stays shallow: one-off commands, high-looking outputs, weakened judgment. Dependency dressed up as adoption. Colombia, Panama, and several Gulf states sit here.
Why it matters
Most talk about the "AI divide" stops at access: who has the tools and who doesn't. That's half the picture. India has low access but its users show strong collaboration rates. Gulf states have high access but directive usage patterns. A country can look "AI-adopted" by every headline metric and still be producing dependency, not empowerment. Policy that only measures adoption misses what happens after adoption.