How AI is Transforming Growth and its Measurement

Modeling AI’s impact on research progress

Benjamin Jones, in this paper, develops a task-based framework to assess how AI capabilities translate into research progress. The model treats research and development as consisting of tasks that can be performed by either machines or humans, with an “ideas production function” mapping R&D inputs to improvements in outcomes. Three critical features determine AI’s impact:

1. The share of research tasks AI can perform
2. The productivity of AI at those tasks relative to humans
3. The strength of bottlenecks in the production process

The framework allows AI to advance along two dimensions—taking over more tasks from humans and becoming substantially better at existing tasks—and derives closed-form solutions showing how these advances affect research progress per dollar of R&D expenditure. The author proposes linking this framework to AI benchmarking studies, where researchers measure AI versus human performance on specific tasks and experimentally identify bottleneck parameters for particular research domains.

Findings

The analysis reveals that bottlenecks severely constrain returns to machine intelligence. (Bottlenecks describe the degree to which progress in research depends on tasks that cannot be easily automated or substantially improved by AI) Even with infinite AI productivity improvements, the rate of progress can only accelerate by limited multiples when AI addresses only a minority of research tasks. For instance, with moderate bottlenecks and an initial capital expenditure share of one-third, infinite machine intelligence would only increase progress rates by a factor of 2.25.

The paper demonstrates two crucial insights:

• Achieving “Transformative AI”—defined as a tenfold acceleration in progress rates—requires substantial automation of human research tasks, typically exceeding 50% of current human tasks, regardless of how intelligent the AI becomes.
• Economically meaningful AI impacts can occur at much lower thresholds than transformative AI, where even modest growth accelerations of 1.2-1.5 times would have profound long-term effects comparable to multiple Industrial Revolutions.

Evidence from production function estimation across hundreds of scientific fields suggests bottleneck parameters are universally negative, supporting the framework’s key predictions.

Artificial Intelligence in Research and Development
Author: Benjamin Jones
From: Northwestern University

Rethinking economic statistics in the age of transformative AI

The paper by Diane Coyle and John Lourenze Poquiz presents a detailed analysis of how transformative AI challenges traditional economic measurement frameworks. The authors organize their methodological approach around the AI value chain, examining measurement issues across inputs, outputs, and impacts of AI:

1. AI inputs: They assess how AI production depends on intangibles like data, computing infrastructure, energy, and specialized labor. The paper draws on national accounting principles, reviews OECD and UN frameworks (SNA25, SEEA EA), and evaluates empirical proxies such as inter-country input-output (ICIO) tables and job vacancy data to estimate AI’s contribution.
2. AI outputs and impacts: The authors use case studies (e.g., healthcare, creative industries) and statistical comparisons (e.g., UK ONS Management and Expectations Survey; DSIT AI adoption survey) to explore how AI-driven service quality and productivity gains are under-measured.
3. Process and time reconfiguration: It employs task-based and time-use analysis frameworks to conceptualize productivity and welfare effects that are not reflected in GDP, focusing on labor-capital complementarities and time savings.

Findings

• Measurement gaps: Conventional GDP and productivity statistics systematically underestimate AI’s contributions because most AI services (especially free ones) are not captured as market transactions, and quality improvements are difficult to deflate accurately.
• Intangible inputs: AI’s development relies on cross-border, intangible, and energy-intensive inputs that current statistical systems fail to allocate correctly across economies.
• Labor and production changes: AI blurs the boundary between capital and labor, potentially lowering the measured labor share while output is misattributed to residual productivity.
• Environmental footprint: AI’s energy use and emissions are inadequately captured by national accounts, requiring multinational coordination of MRIO and environmental satellite accounts.
• Quality and welfare effects: Many welfare gains from AI appear as output-saving innovations—enhancing quality and efficiency while reducing recorded market activity.
• Policy implication: The authors propose a new generation of economic measurement tools to upgrade official statistics through granular task-based data, AI-specific supply-use tables, and time-use indicators to capture how AI transforms productivity and welfare.

Making AI Count: The Next Measurement Frontier
Authors: Diane Coyle, John Lourenze Poquiz
From: University of Cambridge