Answers are grounded in this paper and cite their sources.
Does enterprise adoption of artificial intelligence (AI) raise total factor productivity (TFP), or does measured output growth mainly reflect capital deepening and labour growth? We separate these channels in two stages. First, we compute T\"ornqvist growth-accounting residuals using time-varying labour shares from the Penn World Table (PWT) 10.01, with PWT capital stocks extended to 2024 by perpetual inventory using World Bank investment data. Second, we relate TFP growth to the share of enterprises using AI, taken from Organisation for Economic Co-operation and Development (OECD) business surveys, in a country-year panel with country and year fixed effects. To address endogenous adoption, we use shift-share instruments that interact 2010 fixed-broadband and internet penetration with leave-one-out cross-country AI diffusion, estimated by two-stage least squares (2SLS) and limited-information maximum likelihood (LIML), with weak-instrument-robust Anderson--Rubin inference. Because adoption data cover only 36 countries in recent years (102 country-year observations), statistical power is limited. We find no detectable contemporaneous effect of AI adoption on TFP growth: the fixed-effects estimate is 0.005 percentage points per percentage point of adoption ($p = 0.93$), and the 2SLS and LIML estimates are 0.153 ($p = 0.42$) and 0.299 ($p = 0.48$), with a weak first stage. Point estimates at one- and two-year lags are larger (0.054 and 0.096) but imprecise. A sectoral exposure design with country-by-year and sector-by-year fixed effects, an OECD-only subsample, and a leave-one-country-out jackknife give similar results. The evidence is consistent with, but cannot confirm, a delayed productivity J-curve.
No takes yet. Share an insight, caveat, or question.
Abdullah et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: