Artificial intelligence has been shown to deliver large performance gains in selected economic activities, but its aggregate impact remains debated. This column discusses a micro-to-macro framework to assess the aggregate productivity gains from AI under different scenarios. AI could contribute between 0.25 and 0.6 percentage points to annual total factor productivity growth in the US over the next decade. However, highly uneven sectoral productivity gains could reduce aggregate growth, and large aggregate gains will require a productive integration of AI in a wide range of economic activities.
Artificial intelligence (AI) is transforming what machines can do, from processing natural language to analysing complex datasets and generating images. Recent advances in generative AI (for instance, large language models such as ChatGPT) are also animating a lively debate about the potential for large productivity gains that would allow economies to escape the disappointing productivity growth of the past two decades in many OECD countries (Goldin et al. 2024, Winker et al. 2021, Andre and Gal 2024).
Opinions in this debate vary widely (Figure 1). Some view AI as a transformative general-purpose technology that could unleash productivity growth across a wide range of economic activities and deliver large macroeconomic productivity gains over the next decade (Baily et al. 2023). Others argue that current AI technology is not particularly useful in most economic activities and predict that the aggregate productivity gains from AI will be modest (Acemoglu 2024). Our new paper (Filippucci et al. 2024) contributes to this debate by assessing the aggregate productivity gains from AI under different scenarios for sectoral productivity growth and by discussing the role of sectoral reallocation.
Figure 1 Divergent views about the aggregate productivity gains from AI
Predicted increase in annual labour productivity growth over a 10-year horizon due to AI (in percentage points)
Notes: When the source presents a range of estimates as the main result, the lower and upper bounds are indicated by striped areas. In cases where predictions are made for total factor productivity, predicted labour productivity gains are obtained by assuming a standard long-run multiplier of 1.5 regarding the adjustment of the capital stock (Acemoglu 2024, Aghion and Bunel 2024, Bergeaud 2024 and OECD). The estimates refer to the countries shown in brackets.
Sources: See references at the end of the paper; for Goldman Sachs (2023), the underlying reference is Briggs and Kodnani (2023); for IMF (2024) the underlying reference is Cazzaniga et al. (2024); for OECD, the range from Filippucci et al. (2024) main scenarios are shown (Table 2 last row in Section 3.1).
Sources of disagreement regarding the aggregate productivity gains from AI
A growing body of research documents that AI can significantly increase the performance of workers in specific business contexts, such as customer service (by 14%), business consulting (by 40%), or software development (by more than 50%) (see Filippucci et al. 2024a and 2024b for a review of recent studies on the worker-level productivity impacts of AI). Given the mounting evidence of substantial productivity gains in specific domains, it may be surprising that opinions about the aggregate productivity benefits of AI remain so varied. However, predicting aggregate gains by extrapolating from evidence on the impact of AI in specific parts of the economy is challenging. The economy-wide impact of AI will depend on how broadly AI can be adopted to improve the production processes across many parts of the economy – often referred to as ‘exposure’ to AI – and on how rapidly firms will adopt AI.
In addition, aggregate productivity growth also depends on the relative demand for the goods and services produced in different sectors of the economy. Specifically, a Baumol effect (Baumol 1967, Nordhaus 2008) can arise in general equilibrium if productivity gains from AI are concentrated in a few sectors and relative sectoral demand reacts little to relative price changes. In this case, sectors where AI-driven productivity gains are low (e.g. construction, agriculture, and personal services) may grow as a share of GDP. Aggregate growth could turn out to be limited “not by what we do well but rather by what is essential and yet hard to improve” (Aghion et al. 2019).
We assess the macroeconomic productivity gains from AI under different scenarios for exposure to AI, the speed of AI adoption, and drivers of Baumol’s growth disease. In our main scenarios, we project that AI could contribute between 0.25 and 0.6 percentage points to annual total factor productivity growth in the US (or between 0.4 and 0.9 percentage points to annual labour productivity growth, assuming a standard long-run multiplier of 1.5 regarding the adjustment of the capital stock) over the next decade. Estimates for other economies are of similar magnitude, though somewhat lower, given that adoption of AI is expected to be slower and highly AI-exposed sectors are relatively smaller in these economies.
These predictions, if they indeed materialise, imply a substantial contribution to labour productivity in the context of weak productivity growth across the OECD over the past decades, which has been in the range of 1%–1.5% per year. The upper end of our estimates suggests a productivity gain from AI that is of similarly large magnitude as what has been attributed to ICT in the US during the high-growth decade starting in the mid-90s (around 1% per year; see Byrne et al. 2013 and Bunel et al. 2024).
From micro to macro
To derive projections for macroeconomic productivity growth, we proceed in two steps. First, inspired by Acemoglu (2024), we obtain sectoral productivity gains by combining estimates of worker-level performance gains with measures of sectoral exposure to AI (Figure 2) and projections of future adoption rates based on the historical experience with previous general-purpose technologies (Figure 3). The resulting ten-year sectoral gains in total factor productivity range from 1–2% in manual-intensive activities (agriculture, fishing, mining) to 15–20% in knowledge-intensive services (ICT, finance, professional services), depending on the specific assumptions on AI adoption and exposure.
Figure 2 Exposure to AI varies across sectors
Figure 3 Different scenarios for the adoption path of AI
In the second step, we derive the implied macroeconomic productivity gains using a calibrated multisector general-equilibrium model that accounts for sectoral input-output linkages and the role of demand in driving price adjustments and factor reallocation across sectors (Baqaee and Farhi 2019). Macroeconomic productivity gains are derived under different scenarios regarding the magnitude of micro-level productivity gains, sectoral exposure to AI, the speed of adoption, and structural determinants of sectoral reallocation (Figure 4). The aggregate productivity gains from AI can be decomposed into three effects: (1) a direct effect of increased productivity at the sectoral level; (2) an input-output multiplier effect as productivity gains in one sector also benefit other sectors through reduced costs of intermediate inputs; and (3) a Baumol effect.
Figure 4 Macro-level productivity gains from AI under different scenarios
Estimated impact on annual growth rates of total factor productivity over a 10-year horizon
Notes: The bars correspond to different scenarios regarding the adoption, capabilities, and micro-level gains of AI (as in Figure 1). In scenarios 1 and 2, the elasticity of substitution between the output of different sectors is close to one, and the factors of production (labour and capital) can reallocate freely across sectors. In scenarios 3–5 with adjustment frictions, the elasticity in consumption is assumed to be very low, and factors cannot reallocate across sectors. See more details in section 3 of Filippucci et al. (2024).
AI adoption is a key driver of productivity growth, but uneven sectoral gains could limit aggregate growth through a Baumol effect
A key insight that emerges from this analysis is that the macroeconomic impact of AI will depend primarily on the adoption speed and the degree to which AI can benefit economic activities across a wide range of sectors in the economy. Currently, adoption varies strongly across firms and sectors, with country-level adoption rates being generally low, in the range of 5%–15%, as reported by official statistics of businesses and firm-level studies (e.g. Calvino and Fontanelli 2023a, 2023b). A comparison of scenarios 1 (low adoption) and 2 (high adoption and expanded capabilities) shows that fast and productive integration of AI in a wider range of economic activities through expanded AI capabilities (e.g. further integration with other digital tools) is necessary for the emergence of large macroeconomic gains.
A negative Baumol effect on aggregate productivity growth arises if the productivity benefits of AI are concentrated in a few sectors, as in scenario 3 (high adoption and expanded capabilities, plus uneven sectoral gains and adjustment frictions), where sectoral gains are more uneven because knowledge-intensive sectors such as ICT and finance are assumed to adopt AI more quickly. 1 Productivity gains in the previous technology-driven boom (during the ICT boom decade starting in the mid-90s) were concentrated in a few sectors. In this spirit, scenario 4 (very large gains, concentrated in most exposed sectors, plus adjustment frictions) considers a concentration of sectoral gains that are closer to what was observed during that period. 2 Here, the Baumol effect reduces aggregate productivity gains by a third.
In contrast, no Baumol effect arises if AI gains are more widespread across sectors, for instance if AI is better integrated with robotics technology, which would mean that not only cognitive but also manual-intensive activities could benefit from AI (scenario 5, AI combined with robotics technology, plus adjustment frictions).
We also explore how aggregate productivity effects might depend on the presence of frictions through their impact on changes in the sectoral composition of the economy. Specifically, we consider the possibility that factors of production (capital and labour) cannot be freely reallocated across sectors over our projection horizon. We show that such frictions could magnify the negative Baumol effect by requiring steeper declines in the relative output prices of AI-boosted sectors to create enough demand for their increased output. This would lead to a larger decline in their GDP share, especially if demand is inelastic.Hence, even though such frictions would prevent the reallocation of factors from high- to low-growth sectors, a general equilibrium perspective clarifies that aggregate productivity growth would still be harmed by preventing the efficient allocation of factors towards sectors where they are most valued.
Overall, AI holds promise to revitalise productivity growth in OECD countries and beyond. Governments can also play a role in shaping the macroeconomic gains from AI, for example by resolving legal uncertainties around accountability, which may hold back productive AI adoption by firms (OECD 2024a). At the same time, governments can foster a competitive environment (both in the AI-using as well as the AI-producing sectors; see Aghion and Bunel 2024, OECD 2024b) that is conducive to innovation and experimentation, while monitoring potential labour market disruptions and supporting workers as they transition into new roles in the AI economy (e.g. Acemoglu et al. 2023a,b, Baily et al. 2023, OECD 2023).
Source : VOXeu
