Anthropic claims "Worldwide, uneven adoption remains well-explained by GDP per capita." Using their own data, we show this is not true for middle-income countries—where much of the world's population resides:
Anthropic's Economic Index January 2026 Report claims:
"Worldwide, uneven adoption remains well-explained by GDP per capita."
"At the country level, a 1% increase in GDP per capita is associated with a 0.7% increase in Claude usage per capita."
Using their own data, we show this is not true for middle-income countries—where much of the world's population resides.
Yet I try to show that Anthropic's latest research does not provide much evidence on the question of if AI will create more or less convergence in global income levels between countries. The head of economics at Anthropic, Peter McCrory, told the Financial Times:
"If the productivity gains...materialise in places that have early adoption, you could see a divergence in living standards."
But a cross-sectional regression looking at how income (GDP per capita) impacts AI adoption does not imply divergence in living standards. Impacts on living standards from AI adoption cannot be estimated from this regression. This requires estimating the second order effect of how AI adoption impacts productivity, which Anthropic does not do here.
Moreover, as we show below, our findings suggest more reason to predict convergence in AI adoption at least: middle-income countries are already adopting AI beyond what their income predicts, when allowing for how GDP per capita impacts AI adoption to vary by a country's starting income level.
This sounds like a clean story: richer countries use more AI. But when we analyzed their publicly available data, we found this simple narrative masks crucial complexity.
What if the 0.7 elasticity does not hold everywhere? What if middle-income countries—home to much of the world's population—show a much weaker relationship? That would fundamentally change the policy prescription: these countries don't need to wait to get richer—they can accelerate AI adoption through education, infrastructure, and language access.
That's exactly what we found.
The relationship breaks down mostly for middle-income countries. This matters because middle-income countries contain much of the world's population. For them, income level is a weak predictor of AI adoption—they're adopting AI beyond what their wealth would predict.
The implication: middle-income countries like Brazil, Mexico, Thailand, and Malaysia don't need to wait for more GDP growth in order to get more AI adoption—and aren't. Selective investments in education, digital infrastructure, English proficiency, and regulatory environment may be driving adoption. These are actionable policy levers.
Anthropic also finds that human education—the sophistication of user prompts—correlates with AI adoption. We focus on their GDP per capita claim, which drives the headline, but their education finding supports our argument: middle-income countries can invest in education rather than waiting to get richer to drive AI adoption.
The left panel shows Anthropic's approach: one line through all countries. The right panel shows what happens when we allow the relationship to vary. The slopes are dramatically different.
Note on data coverage: China is not included in Anthropic's dataset. India and Indonesia are included but classified as low-income based on GDP per working-age capita—they are not in the middle-income tercile. The 38 middle-income countries range from South Africa ($9,273 GDP/capita) to Poland ($38,209), and include Brazil, Mexico, Thailand, Malaysia, Colombia, Argentina, Turkey, Chile, Peru, and Romania.
| Income Level | GDP Elasticity (β) | N | What It Means |
|---|---|---|---|
| Low income | 0.76 | 38 | Income strongly linked to AI adoption |
| Middle income | 0.44 | 38 | Adopting beyond what wealth predicts |
| High income | 0.63 | 38 | Moderate relationship, high variation |
| Anthropic (global) | 0.69 | 114 | Masks all this heterogeneity |
Anthropic reports a confidence interval of [0.61, 0.77] — looks precise!
But this ignores that the relationship varies across country groups. When we properly account for this using partial pooling, the interval widens to [0.33, 0.74].
That's nearly 3x wider. Their interval excludes 0.44 as implausible — but 0.44 is exactly what we see in middle-income countries.
Some countries deviate dramatically from the GDP-AI adoption relationship:
Removing these outliers only shifts the global slope by ~5% — so they don't drive Anthropic's estimate. But they highlight that factors beyond GDP per capita — education, language, tech infrastructure, culture — matter for AI adoption.
Higher Income → More AI Adoption
Simple story: richer countries adopt more AI.
One global elasticity applies everywhere.
It Depends on Context
The relationship varies 2x across income levels.
Middle-income countries adopt beyond their wealth.
Income level IS strongly associated with AI adoption. Economic development may be a necessary precondition. Anthropic's story is roughly correct here.
These countries are adopting AI beyond what their income predicts. Brazil, Mexico, Thailand, Malaysia don't need to wait to get richer.
Education, English proficiency, digital literacy, tech infrastructure, and regulatory environment are driving adoption now.
This is where much of the world's population resides — and where Anthropic's simple story breaks down.
Massive variation unexplained by income. Israel is a 3x over-adopter; Gulf states lag despite wealth.
Cultural, linguistic, and policy factors likely dominate. Wealth alone does not predict adoption.
Our most striking finding: middle-income countries are adopting AI beyond what their wealth would predict. Unlike low-income countries (where income level strongly predicts adoption) or high-income countries (where adoption is already high), middle-income nations are finding other pathways to AI adoption.
This is good news: these countries don't need to wait to get richer. Education, infrastructure, and policy can drive adoption now.
All code and data are available at: github.com/IlanStrauss/anthropic-econ-critique
analysis_full.py — Python (statsmodels)analysis_brms.R — R (brms for Bayesian)PARTIAL_POOLING.md — Full technical write-up