AI and the Real Bottlenecks to Growth

Ideas will continue to face barriers to implementation

“A country of geniuses.”

That is Anthropic CEO Dario Amodei’s vision of the world after the arrival of truly powerful artificial intelligence. He foresees millions of independent, highly intelligent agents capable of being sent off to do projects, like an army of smart employees.

Sounds like science fiction, and yet it becomes more plausible by the day. What does economics research have to say about such a world, one where non-human scientists flood the economy with new ideas?

A puzzling disconnect

The typical answer, according to the economic models, is that more researchers leads to more growth.

One complication with this story is that we have already added a ton of researchers over the last few decades, during which productivity growth has been historically weak.

Source: Bloom, Jones, Van Reenen, & Webb (2020)

Economists have been puzzled by these seemingly contradicting trends. To understand why, we need to use just a little bit of math. We promise it won’t be painful.

In 1990, Paul Romer wrote down a model of economic growth that he would eventually win a Nobel Prize for. In that work, Romer structurally incorporated the impact of new ideas on growth.

\(g = \frac{\dot{A}}{A} = \delta H_A\)

This equation says that the growth rate of output and consumption (g) is equal to the growth rate of ideas, which is in turn driven by the number of researchers (H­_A) and how efficient those researchers are (𝛿).

Standard models of economic growth, like the one above, tell us that if you increase the number of researchers, growth should follow.

By pretty much any measure, the amount the US spends on innovation has increased significantly over the past several decades. The share of inventors in the workforce, to take just one indicator, more than quadrupled between 1980 and 2020. Productivity growth has nonetheless lagged.

This is where the growth theory is helpful. If the number of researchers has gone up, but growth remains flat, the models tell us that the efficiency of those researchers must be falling. And indeed this is exactly what one team of economists believes happened.

The cost of new ideas

Economists Nick Bloom, Charles Jones, John van Reenen, and Michael Webb have a worrying conclusion about the ideas economy: “everywhere we look we find that ideas, and the exponential growth they imply, are getting harder to find.”

One place they look for evidence is Moore’s Law. This is the empirical regularity that it takes two years for the number of transistors that fit on a computer chip to double. Moore’s Law has held true for decades, but they found that it takes 18 times more research inputs to keep it going today than it did in 1971.¹

They find similar results across a variety of industries. The amount of research input has gone up, while the real outputs of research have gone down. In other words, researcher efficiency has fallen.

If this is true, this finding has big implications for how we think about AI as an army of researchers. It would mean — again, assuming Bloom and his coauthors are correct — that the main constraint to productivity growth is simply lower gains per researcher, and therefore all we have to do to usher in a new era of rapid productivity growth is to add even more (and potentially better) researchers. A country of more and more geniuses, as it were, is exactly what is called for.

The real constraints to growth

But what if the problem is something else?

New research from Teresa Fort, Nathan Goldschlag, Jack Liang, Peter K. Schott, and Nikolas Zolas calls into question the finding that ideas are getting harder to find. Looking at more than four decades of firm-level data on research investments, patents, and the growth outcomes of all non-farm employer businesses, they break down the relationship between ideas and economic growth into two parts: (1) the relationship between R&D investments and ideas generated, and (2) the relationship between ideas and firm growth.

Fort and coauthors find:

1. R&D→Ideas: Ideas are NOT getting harder to find. If anything, firms are getting more ideas per dollar spent on research.

2. Ideas→Firm Growth: Ideas are NOT getting harder to translate into firm growth. The ability of firms to translate ideas into growth is strong and relatively stable over time.

But we are now left with a mystery. If the returns to R&D aren’t falling, and the relationship between ideas and growth is stable, then why hasn’t the significant rise of investment in innovation led to booming growth?

The key turns out to be secular trends in firm growth — patterns that are common to all firms after accounting for ideas. Fort and her coauthors find that these secular trends, which are experienced by all firms equally, are pushing down firm growth rates. Which means that we need to focus our attention on other trends in the economy as the culprit for why increased research expenditures are not translating into growth.

AI and the Bottleneck(s) to Growth

Even if AI ends up helicoptering millions of new geniuses and their ideas down onto the economy, the work of Fort and coauthors suggests it might not automatically translate into a productivity boom. If researcher efficiency wasn’t the bottleneck, we need to place more focus on the secular trends occurring across the whole economy that are dragging down firm growth rates.

What might these secular trends be? That question falls outside the scope of their paper, but we can think of a few candidates.

1. Diffusion of Ideas: Ufuk Akcigit and Sina Ates show that the slowing diffusion of ideas from the most productive firms to others is consistent with a broad set of macroeconomic trends including rising market concentration, increased average markups, declining firm entry, and decreased dispersion in firm growth.

2. Demographic Change: The path of dynamism is intimately linked to changing demographics. Fatih Karahan, Benjamin Pugsley, and Ayşegül Şahin show that the slowdown in labor supply growth since the 1970s was an important driver of the long-run decline in the startup rate.

3. Regulation and Rent-seeking: Large incumbent firms often attempt to leverage state power to gain advantage and stifle competitive pressure. Even though Nathan Goldschlag and Alex Tabarrok find that federal regulation alone cannot explain declining dynamism, Ufuk Akcigit, Salomé Baslandze, and Francesca Lotti show that market leaders are much more likely to have political connections and much less likely to innovate. Mancur Olson provides a more sweeping account of collective action problems and how concentrated interest groups slow growth.

One common thread from these other factors is that our productivity problems are not about how much innovation is happening at one particular firm, but how that innovation does or doesn’t spread throughout the economy.

This possibility is supported by research showing declining skewness in firm growth rates and declining responsiveness of firms to shocks. In plainer language, the idea is that we have fewer high-growth firms, and also that those firms that come up with a new product or process, or see a surge in demand for some other reason, on average don’t create as many jobs as they used to. Complementary research shows increased concentration among market leaders, with market leaders increasingly pulling away from their competitors in how productive they are (in economics jargon, there is rising skewness in productivity between firms). Firms on the frontier appear to be doing fine, but laggards are falling further behind market leaders over time.

Improving the allocation of funding for science is an important and necessary policy goal. So is future-proofing our institutions, readying them for a higher rate of discovery. But ideas alone might not be the economic sparks we hope them to be if they don’t lead to fires that spread throughout the economy.

This is not to say that AI won’t increase growth through other channels, such as reducing the cost of providing certain goods or services, or even that the ideas we get from AI won’t increase growth. Instead, we are suggesting that there will remain many important bottlenecks.

Researchers and policymakers need to place more emphasis on understanding and addressing a broader set of drags on growth — perhaps with AI! — that go beyond the Eureka moment. We need to reinvigorate entrepreneurship, economic churn, and the flow of labor, capital, and ideas across firms. In short, they need to pursue policies that increase America’s economic dynamism — or risk squandering the talent of our future AI geniuses.

1. They measure research input by taking total research expenditures divided by the average cost of a skilled worker. This implicitly includes both capital and labor expenditures, but they argue this approach captures the “effective number of researchers.” They find the result is consistent across different measurement choices.