This article synthesizes a literature review to explore whether artificial intelligence can drive productivity growth.
What type of technological progress does AI represent?
Historical technological advances have varied in type and impact on productivity. The academic consensus holds that AI exhibits the characteristics of a General Purpose Technology (GPT), capable of boosting productivity over long cycles. Empirical data already shows an uptick in US labor productivity since 2023.
Can AI deliver sustained long-term productivity gains?
Two main schools of thought exist. One, exemplified by Gordon, argues that technologies emerging after the Second Industrial Revolution generally fail to deliver lasting productivity improvements. The other, represented by concepts like the "Solow Productivity Paradox" and the J-curve effect, posits that information technology and AI, as GPTs, can enhance long-term productivity, though their impact may be muted in the early stages due to time lags, resource misallocation, and measurement errors.
When will AI likely boost productivity?
Current research broadly predicts the most significant productivity lift from AI will occur around the 2030s.
How does the "AI Productivity Paradox" affect monetary policy?
Scholars generally agree that central banks should pay closer attention to structural economic shifts during transitions and place greater weight on micro-level business survey data, which often signals changes earlier than aggregate statistics. Monetary policy is advised to remain relatively accommodative to support the diffusion of new technologies.
Summary of Findings
1. The Nature of AI as a Technology
Historical progress includes distinct types like "lightbulb" technologies, "electric motor"-type GPTs, and "microscope"-type "inventions of method." AI possesses dual attributes of both a GPT and an "invention of method." On one hand, it shows classic GPT traits: widespread diffusion across economic sectors, continuous improvement over time, and the capacity to spur further innovation. On the other, as an "invention of method," it can optimize organizational structures, enhance R&D efficiency, and spawn new business models.
US labor productivity has already risen since 2023. In aggregate terms, it has significantly exceeded its pre-pandemic trend. A decomposition of productivity contributions reveals that the recent gains have been primarily driven by capital deepening (increasing from 0.3 percentage points to 0.9 pp) and improvements in labor quality (rising from 0.2 pp to 0.4 pp), while total factor productivity growth has remained more modest.
2. The Debate on Long-Term Productivity Impact
The question of whether post-internet technologies can durably raise productivity is contested. Gordon's seminal 2000 work contrasted the transformative, life-altering impact of the Second Industrial Revolution (electricity, internal combustion, etc.) with the more limited effect of the information revolution, which primarily influenced entertainment and communication, arguing the latter cannot sustain high long-term productivity growth. Some contemporary research echoes this skepticism regarding AI, citing plateauing education levels, the internet's broader initial penetration, and AI's focus on "information" rather than physical "energy."
The opposing view centers on the "productivity paradox." This phenomenon, famously noted by Robert Solow, describes the disconnect between rapid IT investment and stagnant measured productivity in the 1980s/90s. Explanations include time lags, measurement issues, and resource misallocation. Current analysis applies this lens to AI, identifying several key factors.
The Time Lag Hypothesis
GPTs like steam, electricity, and semiconductors require complementary organizational changes, human capital development, and broad diffusion, leading to long gestation periods for productivity dividends. The "productivity J-curve" quantifies this, suggesting early investment in intangible assets (new processes, business models) depresses measured output before later yielding gains. Research indicates AI follows this pattern. Micro-level studies, such as one on US manufacturing, found initial AI adoption lowered firm-level productivity, but firms persisting with AI from 2017 to 2021 showed significantly faster productivity growth later.
Resource Misallocation and Concentration
The benefits of new technology often concentrate in "superstar" firms, masking aggregate productivity gains. Studies show AI's productivity impact is highly uneven across firms and sectors. Research indicates current productivity rebounds are not broad-based but concentrated in specific sub-sectors like data processing and computer systems design. A 2026 multi-central bank survey found AI adoption is significantly higher in larger, more productive firms and in sectors like finance and IT, leading to pronounced industry divergence.
Measurement Errors
National accounting methods, slow to adapt to new technologies, can create statistical errors in early stages. Current frameworks may undervalue intangible assets and AI services. For instance, when firms use AI via API/cloud subscriptions, these costs are often treated as intermediate consumption rather than capital investment, potentially distorting productivity calculations by overstating intermediate inputs and understating capital. However, most research suggests measurement error, while present, is not the primary explanation for the paradox.
3. The Timeline for Productivity Gains
Historical analysis suggests lags between GPT emergence and productivity impact range from 10 to 40 years, with the lag shortening from steam to IT. Current predictions commonly point to the 2030s for AI's most significant effect. One model outlines three phases: a "paradox" period (2022-2025) with high investment but little measured gain; an "impact emergence" phase (2026-2029) where benefits appear in leading firms and sectors; and a "productivity acceleration" stage (2030-2035) with widespread diffusion and measurable macro gains. Different sectors are expected to reach acceleration at different times, from fast-diffusing industries like IT by 2028-29 to slower sectors like traditional manufacturing after 2035.
4. Implications for Monetary Policy
Given the "high input, low output" phase associated with the paradox, research explores monetary policy implications. As the first two industrial revolutions occurred under the gold standard, analysis often draws on the Greenspan era during the IT revolution. A common conclusion is that central banks should focus more on structural changes and micro-level survey data for early signals. Policy should maintain relative accommodation to support technology diffusion.
Research suggests AI may create a "two-speed economy," with inflation in AI-expanding sectors representing structural adjustment rather than demand overheating, warranting policy tolerance. Policymakers must distinguish between cyclical weakness and structural dislocation caused by technological shifts. The experience of the 1990s shows that allowing growth to "run" during a period of mild inflation can facilitate a smoother transition, though it requires incorporating macroprudential tools to address financial stability risks from prolonged accommodation. Central banks are advised to incorporate real-time business insights while avoiding over-optimism that could fuel investment bubbles.
Comments