> This paper provides the first multi-wave, cross-country evidence on both the aggregate and distributional consequences of AI adoption. Using five releases of the Anthropic Economic Index spanning January 2025 to February 2026, we construct two novel measures from observed AI usage data: a labor cost equivalent (LCE) estimate of AI’s aggregate productivity value, and an AI concentration index (ACI) that tracks whether gains flow disproportionately to higher-paid or lower-paid occupations.
> Aggregate AI gains are large and rising. We estimate approximately $2.7 trillion in annualized labor cost equivalent at own-country ILO wages, and the figure is increasing as AI enters occupations that employ far more workers, even as the per-conversation value of AI declines. Income and regulatory readiness determine the starting point: richer countries and countries with higher regulatory readiness have lower ACI, and regulatory readiness is the dominant predictor of LCE relative to GDP. But neither predicts the trajectory; faster diffusion is associated with official English- language status, consistent with the interpretation that countries whose institutional knowledge is well represented in the training data of English-dominant large language models benefit from more accurate, locally relevant AI assistance across a wider range of occupations.
1 Comment
> This paper provides the first multi-wave, cross-country evidence on both the aggregate and distributional consequences of AI adoption. Using five releases of the Anthropic Economic Index spanning January 2025 to February 2026, we construct two novel measures from observed AI usage data: a labor cost equivalent (LCE) estimate of AI’s aggregate productivity value, and an AI concentration index (ACI) that tracks whether gains flow disproportionately to higher-paid or lower-paid occupations.
> Aggregate AI gains are large and rising. We estimate approximately $2.7 trillion in annualized labor cost equivalent at own-country ILO wages, and the figure is increasing as AI enters occupations that employ far more workers, even as the per-conversation value of AI declines. Income and regulatory readiness determine the starting point: richer countries and countries with higher regulatory readiness have lower ACI, and regulatory readiness is the dominant predictor of LCE relative to GDP. But neither predicts the trajectory; faster diffusion is associated with official English- language status, consistent with the interpretation that countries whose institutional knowledge is well represented in the training data of English-dominant large language models benefit from more accurate, locally relevant AI assistance across a wider range of occupations.