Essays on growth accelerations, knowledge transmission, and the hidden costs of Artificial Intelligence

Chapter 1 investigates the determinants of growth accelerations in GDP per capita across 148 countries since the 1950s. Growth accelerations are identified using the filter introduced by Hausmann et al. (2005), and the chapter revisits the econometric framework of Koopman and Wacker (2023). We show that, once the time-series properties of the data are properly accounted for, physical capital accumulation becomes a statistically significant predictor of growth accelerations, but only in capital-abundant economies. The chapter establishes an empirical regularity—namely, that the growth payoff of capital accumulation is conditional on the existing stock of capital—which motivates the theoretical and empirical study in Chapter 2. Chapter 2 develops a multi-country AK model of endogenous growth with international knowledge transmission to study the role of productive public expenditure in growth and convergence. The framework combines the productive-public-spending mechanism of Barro (1990) with the learning by-investing externality of Romer (1986) in a leader–follower setting. A single frontier economy advances the world technology frontier, while follower economies benefit from international spillovers only after crossing a critical absorptive-capacity threshold, measured as capital per capita relative to the leader. Below this threshold, productive public expenditure is the main engine of growth; once the threshold is crossed, knowledge transmission becomes operative and follower growth converges toward the leader’s rate. On the empirical side, using dynamic heterogeneous panel models for 110 countries over 1970–2019, the chapter identifies a statistically significant absorptive-capacity threshold at about 20% of U.S. capital per capita. Below the threshold, productive public spending is associated with higher long-run growth, while spillovers remain limited. Above the threshold, spillovers become the main force behind convergence, with effects that gradually attenuate as the technology gap narrows. These findings are robust to richer empirical specifications and are further reinforced when technology diffusion is proxied through capital imports: this channel provides long-run growth effects only for countries that have already crossed the threshold. Chapter 3 takes a different perspective by moving to the technological frontier and examining a tension that is largely absent from existing AI-and-economic models: broad AI adoption may homogenize the product space and partly offset the productivity gains it delivers. The chapter develops a general-equilibrium model of monopolistic competition with heterogeneous final-good firms in the tradition of Melitz (2003), together with two intermediate sectors supplying AI and non-AI services. Each final-good firm produces a single differentiated variety and draws an exogenous capability parameter. Higher capability raises product quality, but higher-quality varieties are more costly to produce. AI affects welfare through two forces. First, higher AI productivity lowers the fixed organizational cost of market participation. This expands the mass of active varieties, generating a love-of-variety gain, but also lowers the survival cutoff, weakens selection, and allows lower-capability firms to operate. Second, AI generates a monoculture externality. As firms rely more on common AI inputs, average AI exposure rises and the market-relevant distinctiveness of final-good varieties declines. Because each firm takes average AI exposure as given, the decentralized equilibrium does not internalize this product-space externality. Drawing on recent experimental evidence that large language models can improve individual outputs while reducing diversity across producers Doshi and Hauser (2024), Meincke et al. (2025), the chapter characterizes the conditions under which AI productivity gains dominate the losses from weaker selection and lower distinctiveness, and the conditions under which the opposite is true. The welfare analysis shows that AI is not unconditionally beneficial: a corrective AI tax is optimal when the monoculture channel dominates, whereas an AI subsidy is optimal when organizational-cost gains dominate.