S9 Ep18: Will AI transform economic growth?

Could AI transform our economies to produce explosive growth? Most economists are sceptical at best. Anton Korinek of the University of Virginia, leader of the CEPR research policy network on AI, thinks the threshold is closer than those models suggest.

In his latest work, Korinek, Tom Davidson, Basil Halperin, and Thomas Houlden, have built a growth model that captures what happens when AI starts automating AI research itself. Automation does two things simultaneously: it accelerates research, and it offsets the diminishing returns that have historically stopped self-improving processes from compounding. Three reinforcing feedback loops: software quality, hardware quality, and general technological progress, each amplify the others. Korinek's findings are more optimistic than even the AI labs' own roadmaps, which focus on software capability alone.

The research behind this episode:

Davidson, Tom, Basil Halperin, Thomas Houlden, and Anton Korinek. 2026. "When Does Automating AI Research Produce Explosive Growth? Feedback Loops in Innovation Networks." Working paper, January 2026.

To cite this episode:

Phillips, Tim, and Anton Korinek. 2026. "When Does Automating AI Research Produce Explosive Growth?" VoxTalks Economics (podcast).

Assign this as extra listening. The citation above is formatted and ready for a reading list or VLE.

About the guests

Anton Korinek is a professor of economics at the University of Virginia. He leads the CEPR Research Policy Network on AI, which is building a community of researchers to understand and anticipate the economic impact of artificial intelligence. He is a member of Anthropic's Economic Advisory Council and was named by Time magazine among the hundred most influential people in AI. His research spanning the economics of transformative AI, growth theory, and the implications of advanced automation for labor markets and inequality has made him one of the most widely cited economists working on these questions. He is also the founder of the Economics of Transformative AI initiative at the University of Virginia, which focuses on the long-run economic consequences of AI systems that approach or exceed human-level capabilities.

Visit the CEPR Research Policy Network on AI.

Research cited in this episode

Daron Acemoglu's estimate of AI's growth impact. Acemoglu calculated that AI would raise annual growth by approximately 0.07 percentage points, arriving at this figure by multiplying the share of jobs likely to be affected by AI, the fraction of tasks within those jobs that AI could perform, and the productivity gain per task. Korinek argues the estimate was a reasonable description of the AI that existed in 2024 but did not account for the trajectory of capabilities since, nor for the feedback loops between AI progress and further AI development that his own paper models.

Recursive self-improvement. The idea that an AI system, once capable enough, could design improved versions of itself, triggering an accelerating cycle of capability gains. The concept was first articulated by John von Neumann in the 1950s and has since become central to debates about transformative AI. All major AI labs, Korinek notes, are working towards some version of this vision; the economic question is whether the resulting growth would be explosive or would be damped by diminishing returns.

Semi-endogenous growth models. A class of economic growth models in which long-run growth depends on the scale of the research workforce and the returns to research effort. The canonical insight, associated most closely with Nicholas Bloom and co-authors, is that "ideas get harder to find"; maintaining a given rate of progress requires ever-increasing research investment. Korinek and co-authors use and extend this framework, showing that automation can counteract diminishing returns by replacing human labor with capital in the research process, creating a new feedback loop that was absent from earlier models.

Kaldor's balanced growth facts. Nicholas Kaldor's observation, made in the mid-twentieth century, that the major macroeconomic aggregates, including the capital-output ratio, the labor share of income, and the rate of return to capital, remain roughly stable over long periods. Growth economists built their models, including the Solow and Ramsey models, to fit these regularities. Korinek notes that those models were appropriate precisely because they matched the historical data; the question his paper raises is whether the data of the next few decades will look different enough to require a different class of models.

Moore's Law. The empirical regularity, observed in computing hardware since the 1960s, that the number of transistors on a chip approximately doubles every two years. Korinek uses chip progress as a calibration benchmark: maintaining that rate of doubling has historically required roughly an eight percent annual increase in the scientific workforce working on chips. This figure allows the model to be parameterised with a real-world measurement of how much additional research input is needed to sustain a given rate of technological progress.

Consumer surplus from digital technologies. Korinek raises the problem that GDP statistics are designed to measure market transactions and therefore do not capture the value people derive from digital goods and services beyond what they pay for them. He references research from the Stanford Digital Economy Lab as an example of work attempting to quantify this surplus. The implication for the paper's argument is that explosive AI-driven growth could be underestimated even in the statistics used to monitor it.

More VoxTalks Economics episodes

"Our Workless Future", an earlier conversation with Anton Korinek from September 2022, in which he set out the case for taking AI's impact on labor markets seriously.

Related reading on VoxEU

Firms predict an AI productivity boom is coming, a survey of over 5,000 CFOs, CEOs, and executives shows that around 70% of firms actively use AI, particularly younger, more productive firms. They forecast AI will boost productivity by 1.4%, increase output by 0.8%, and cut employment by 0.7% over the next three years.

How AI is affecting productivity and jobs in Europe, firm-level evidence on AI’s effects in Europe. The authors find that AI adoption increases labour productivity levels by 4% on average in the EU, with no evidence of reduced employment in the short run.

From AI investment to GDP growth: An ecosystem view, how the current AI wave is contributing to US GDP, both directly through investment and indirectly through ongoing service flows.