The case for—and against—rapid AI-driven growth

I talked to several experts about how AI might transform the economy.

A big topic of conversation at this year’s World Economic Forum in Davos was the potential for AI to boost the economy. Bill Gates predicted that AI is “going to raise productivity generally.” He said that AI is already having a “dramatic” impact on the productivity of white-collar workers, and he predicted that robotics would eventually increase blue collar productivity as well.

"I'm actually very bullish about some of the possibilities for significantly improving productivity much faster than we might have thought," said economist Erik Brynjolfsson.

How much could AI boost the economy? In 2021 Open Philanthropy, a leading funder of effective altruist causes, published a paper that tried to answer that question. The author, Tom Davidson, came up with a startling prediction: annual economic growth could soar to 30 percent before the end of the century.

That’s about 10 times faster than typical growth over the last century. It would mean the economy more than doubles every three years and the world would get 1,000 times richer in a generation.

But Ben Jones, a Northwestern University economics professor who Open Philanthropy recruited to review Davidson’s paper, argued that Davidson’s figure was implausible.

“The world is full of amazing technologies, including those with wide ranging applications,” Jones wrote. “But the better they get, the more we are left struggling with what those technologies don’t do.”

In a November post, economist Samuel Hammond argued that “the herculean task of maintaining ~3% annual GDP growth in the context of a frontier economy should put some strain on the plausibility of growth rates running substantially higher.”

There are actually two questions here. One question is “given human-level AI, how fast will the economy grow?” As I talked to experts on both sides of this debate, I found there was actually less disagreement than I expected. While many are skeptical of Davidson’s 30 percent figure, there is broad agreement that human-level AI would have a big economic impact.

The other key question is “how likely are we to get human-level AI in the next few decades?” And this turns crucially on what we mean by human-level AI.

Davidson’s model requires AI systems to be sophisticated enough to replace human workers across the economy. That not only requires AI systems with human-level cognitive abilities, it also requires robots capable of performing a wide range of physical tasks, from fixing a leaking toilet to taking care of toddlers.

Creating robots with physical abilities like this could prove just as difficult as creating software with human-level cognitive abilities. And even if we manage it, human consumers may still prefer to have other human beings provide many face-to-face services. That could constrain growth rates and leave plenty of space for human workers to earn a living no matter how sophisticated technology gets.

Setting the stage

Those who predict a future of explosive growth note that we’re already starting to see signs of AI’s economic potential. Many white-collar workers are using generative AI technology, and some studies have shown sizable productivity gains. These results are especially impressive considering how nascent these tools are. 

As the technology becomes more advanced, it will enable AI research itself to become more productive. Powerful AI systems could help human researchers design still more powerful AI. AI systems might start to improve their own programming with little to no human assistance. 

Davidson expects human-level AI will allow us to put billions of virtual AI researchers to work improving AI systems. This positive feedback loop could lead to an incredibly fast period of innovation and enable AI systems to automate nearly all cognitive tasks, from therapy to the practice of law.

While Davidson sees a 50 percent chance of  human-level AI by 2040, others expect it to take longer. For example, I spoke to Ege Erdil, a researcher at Epoch, an AI-focused think tank funded by Open Philanthropy. He believes it will take about 40 years to develop AI capable of automating all remote jobs.

While that would significantly increase economic output, it probably wouldn't cause explosive growth on its own. That’s because most jobs require taking actions in the physical world. And here the case for explosive growth is—if not weaker—at least vaguer.

Those predicting explosive growth assume that cognitive advancements will trickle into the physical world. But they tend not to go into a lot of detail about how this will work.

The basic idea is that the invention of human-level AI systems will allow us to massively increase the amount of brainpower devoted to robotics research. This implicitly assumes that intelligence, rather than real-world experimentation, is the main bottleneck to progress in the robotics field—an assumption Tim critiqued a few weeks ago. But let’s assume advanced AI drives much faster progress in robotics.

Many jobs might still prove highly resistant to automation. For example, no matter how sophisticated AI software or robotic hardware gets, it’s hard to imagine a parent entrusting their child to a robot that looks like the Atlas Robot from Boston Dynamics:

Photo via Boston Dynamics

And this wouldn’t be misguided anti-robot prejudice. Love and affection are essential to the healthy development of young children, who are unlikely to form a healthy relationship with a robot no matter how competent it becomes at the mechanical aspects of the job.

Or consider coffee shops. We already have machines that are capable of producing an excellent cup of coffee. Yet many people remain willing to pay for another human being to operate the machine on their behalf. Even if we had robots capable of doing exactly the same tasks as a human barista, it’s likely that many customers would choose to patronize coffee shops with a human being behind the counter.

The case for explosive growth

Even if AI struggles to replace the most sensitive care industries, it seems plausible that AI and robotics could take over a large number of jobs, including in factories. We could wind up with a virtuous cycle in which robots help us build more and better robots.

Jones is skeptical of Davidson’s 30 percent growth figure, but in a phone interview he acknowledged that robots that can physically and mentally perform at a human level could produce “really rapid growth and higher standards of living.”

Basil Halperin, an MIT PhD student in economics, told me that people would likely choose to work much less in a world of advanced AI because everything would be so cheap. And although inequality could become incredibly stratified according to who owned the machines, Erdil said “it will be better to be a poor person in this world than be even a rich person today.”

So how much could this virtuous circle ultimately accelerate growth? Davidson’s paper points back to the Industrial Revolution, when typical growth rates increased from 0.3 percent to 3 percent:

Davidson believes the AI revolution will produce another order-of-magnitude increase in the annual growth rate, from 3 percent to 30 percent, as we let robots take over a larger and larger share of economic activity. But even assuming rapid progress in AI and robotics, there may be inherent limits to the speed of economic growth.

Sectoral bottlenecks

The first problem is a phenomenon economists call Baumol's Cost Disease. As the economy gets more productive, relatively less efficient sectors make up a larger share of the economy.

Think about agriculture. The world is producing more crops and doing so more efficiently than ever. As a result, we spend less on food production, and so the sector makes up a shrinking share of the economy. At this point, getting even better at corn production won’t move the needle much on GDP.

We could see a similar phenomenon as AI automates some, but not all, sectors of the economy.

“When you get really good at something,” Jones told me, “what you end up doing is focusing on the things you're less good at.” He pointed to the service sector which has had less productivity growth relative to others and, as a result, now dominates the American economy.

The reason for this is simple: consumers only want so much of any single good or service. In the 19th century, the average family only had one or two changes of clothes per person. So people bought more clothing as it got cheaper and the textile industry grew rapidly. But by the mid-20th Century, most Americans had closets full of clothing. So as the cost of clothing continued to fall, we just spent a smaller share of our incomes on apparel. We poured the savings into labor-intensive services like health care, education, and recreation instead.

This basic story of technological advancement amidst slower growth in human productivity has arguably been the story of the past several centuries and could continue in a world of AI, according to University of Virginia Professor Anton Korinek. He told me that “if there are some goods that only humans could produce (say, human connection)—and of course assuming that humans remain in control—then those goods would become relatively more expensive,” the same way labor-intensive services have become more expensive relative to mass-produced consumer goods.

So, human-level AI may very well make many service jobs more efficient. But that would cause consumer spending to increasingly shift to harder-to-automate sectors, and as a result these sectors will increasingly determine the economy-wide growth rate.

To be fair, there are some spillovers between sectoral growths in productivity. Economists call this the “elasticity of substitution between labor and capital.”  If the minimum wage goes up too much, for example, companies switch to self-checkout machines. If robots get cheap enough, they may replace human labor even if they aren’t complete substitutes. For this reason, Erdil told me that “Even if you cannot do some of the tasks of the economy, you can still get explosive growth.” 

This fundamental question — to what extent capital and labor are interchangeable — is hotly debated by economists. Its answer is incredibly important for whether transformative growth is possible. If AI advances are siloed in a given job or sector, automation can be very successful without adding all that much to economy-wide growth. Indeed, in a back and forth with Jones, Davidson admits that “bottlenecks will eventually slow growth,” yet maintains that there could be an interim period of explosive growth.

Fundamental limits

It’s not just that some sectors will be harder to automate than others. There are also fundamental limits that can’t be programmed away, a point that economists often bring up when criticizing predictions of extreme AI-driven growth.

For starters, the laws of physics sometimes limit how much technology can improve. “You can’t have more efficient energy forever,” Jones said, “because you run into the second law of thermodynamics.”

This may seem like an extreme upper bound, but remember how fast 30 percent annual growth would be. Thousands of years of progress would be crammed into a few decades as our knowledge compounds over and over again. Questions would arise very quickly surrounding how much better our usage of energy, land, or natural resources can get.

If one assumes that transformative growth only occurs briefly and not for an extended time, these criticisms become much less persuasive since these fundamental limits likely won’t bind during a brief expansion. Erdil told me that “most of the land on Earth is actually not used” and humans only use “0.01 percent of the energy that Earth gets from the Sun.”

Economies will hit many limits before reaching those dictated by the laws of physics. Take self-driving cars. Even if every car was autonomously piloted, Jones told me, you can only get from point A to point B so fast due to congestion concerns. Samuel Hammond makes a similar point in a recent post, noting that you “can’t cure cancer twice.”

Erdil pointed out to me that even at current growth rates, we’ll hit these limits eventually. So perhaps explosive growth should be understood as taking us to the economic frontier much faster before a subsequent period of stagnation.

Korinek offered another helpful way to conceptualize transformative growth amidst these barriers. He told me that in a world of very advanced AI, “the growth we are likely to experience is that much of it would be about quality improvements and new goods rather than just quantity.”

One example of this principle Erdil gave me was that even if you can’t make planes much faster, you can make the flying process much more comfortable. So, we may not have double the number of buildings every three years, but our homes will be much nicer. Still, fundamental limits might still apply, as there may be a ceiling on how much more comfortable a home or airplane ride can get.

The form these improvements could take is incredibly uncertain. “This is all extremely speculative,” Korinek told me. “The kind of economy we are discussing is so different from our current economy that it is really hard to say anything with certainty.”

Ultimately, growth may be limited by demand as much as by supply. There’s no doubt that robots and AI will expand our capacity to make more stuff. But how much stuff we make will depend on how much stuff consumers want to buy. And as we collectively get wealthier, we may find that what we value most are personal connections that can only be supplied by other people.

If you enjoyed this piece please follow Aden on Twitter.