Generative AI as an Inequality Amplifier: Language Compatibility, Leapfrogging and Productivity Growth

We provide the first causal evidence that generative AI amplifies global inequality through linguistic compatibility with training data. Exploiting the November 2022 release of ChatGPT as a natural experiment across 169 economies, we find that a onestandard-deviation increase in Language Intensity-a country's linguistic compatibility with AI training data-raises productivity growth by 0.94 percentage points. This produces "linguistic leapfrogging": low-income but linguistically compatible economies-Ghana, Nigeria-match the gains of wealthy Anglophone nations, while high-income but incompatible economies like Japan see none. We conduct a controlled experiment to provide micro-foundations: across twelve languages, a generative AI tool delivers business decisions of lower economic value and higher cost as language resourcedness falls, and while a frontier model nearly closes the quality gap, the cost gap-fixed at the tokenizer-is structural and does not. A structural model of AI-augmented technology adoption, calibrated to our empirical estimate, projects these patterns forward and shows that multilingual AI development would cut the inequality cost of AI-driven growth by 59 percent. Our results challenge standard views of technology diffusion and reveal that AI introduces a new axis of global stratification along linguistic lines.