An illustration of a bright coral-colored factory room filled with three rows of blue and green robots doing manual work with wrenches.
Caleb Luke Lin for Vox

And how it could fizzle.

Artificial intelligence is already making people rich. Jensen Huang, the co-founder and CEO of chip company Nvidia, which controls 80 percent of the data-center AI chip market, has seen his net worth explode from a mere $4 billion five years ago to a staggering $83.1 billion as of March 24 on the back of bottomless demand for his company’s product.

ChatGPT maker OpenAI is reportedly valued at $86 billion, with rivals Anthropic and Inflection at $15 billion and $4 billion as of their most recent funding rounds. While OpenAI CEO Sam Altman says he owns no shares in the company, it’s possible, even likely, that other AI founders and execs have joined the three commas club by now, at least on paper.

But some researchers think this is only the beginning — that AI won’t just make a few techies wildly rich, the way social networking, smartphones, and personal computers did before. Believers in a growth explosion argue that AI is set to make society much, much richer by causing economic growth at a scale it has never experienced before.

In 2020, the AI researcher Ajeya Cotra at grant maker Open Philanthropy released a report arguing that AI powerful enough to drive a surge in economic growth to 20 to 30 percent a year is coming, and more likely than not will emerge before 2100. The following year, her colleague Tom Davidson conducted a more in-depth investigation of the potential for AI to supercharge growth and concluded that per capita economic growth rates as high as 30 percent a year resulting from AI are plausible this century.

This is an extremely “big if true” claim. Since good record-keeping began shortly after World War II, the US has averaged 3.2 percent economic growth per year. Since 2000, growth has been much more anemic, averaging 2.2 percent. Per capita growth — which is affected by population changes as well as economic ones — has been lower still.

Nowhere before in history — not in England during the Industrial Revolution, not in Japan during its “income doubling” period in the 1960s, not in China in recent decades — has sustained growth on the scale of 20 to 30 percent per year happened. To put that number into perspective, 30 percent growth implies that the economy would double in size every 2.5 years or so. (Based on current growth levels, the US economy won’t double for 35 years.)

It gets even more impressive when you take a longer view. Northwestern economist Ben Jones has noted the typical American today is about 100 times richer than humans were when economic growth began and we were all living at the edge of starvation. In a system of 30 percent growth per capita, in 25 years we’d be 1,000 times richer than we are now.

Imagine everything humans have achieved since the days when we lived in caves: wheels, writing, bronze and iron smelting, pyramids and the Great Wall, ocean-traversing ships, mechanical reaping, railroads, telegraphy, electricity, photography, film, recorded music, laundry machines, television, the internet, cellphones. Now imagine accomplishing 10 times all that — in just a quarter century.

This is a very, very, very strange world we’re contemplating. It’s strange enough that it’s fair to wonder whether it’s even possible. Personally, 30 percent growth is so far outside human experience to date that I have trouble even imagining what it might look like.

AI could be just another useful technology, akin to a washing machine. In this view, it makes our lives a little better, like most technological improvements.

But AI could also be something else entirely that would upend the assumptions we’ve used to understand the world around us for centuries.

The basic case for a growth explosion

In his 2021 report, Davidson lays out three general arguments for why such a dramatic explosion in economic growth might be possible.

The first argument is historic. In an earlier report for Open Philanthropy, researcher David Roodman looked at the trajectory of the world economy in the very, very long run — all the way back to 10,000 BCE. He concluded that the pattern of economic growth, examined through this very wide lens, is superexponential. Exponential growth means the economy grows by a steady, compounding rate every year — 2 or 3 percent, say — like interest in your savings account. Superexponential growth means that the growth rate is increasing over time. That, Roodman concludes, is what has in fact happened.

Roodman emphasizes that you should take this with several grains of salt. It’s not like we have good data on what the world economy was like in 10,000 BCE. But we do know, with a high degree of confidence, that economic growth was very slow for a very long time and then accelerated a great deal with the onset of the Industrial Revolution.

That fits a superexponential story. And a superexponential story makes future increases in the rate of economic growth look very plausible. “Some people have the prior of ‘This is crazy’” when thinking about superexponential growth, Davidson told me in an interview. “And other people have the prior of ‘This has happened throughout history.’”

Davidson’s second argument relies on a popular set of theories within economics for why growth has accelerated over the very long run. The short answer these theories give is that population growth enabled economic growth to speed up.

“A long time ago, the world population was relatively small and the productivity of this population at producing ideas was extremely low,” Stanford economist Chad Jones explains in a 2001 paper. “Once an idea was discovered, however, consumption and fertility rose, producing a rise in population growth. More people were then available to find new ideas, and the next new idea was discovered more quickly.”

Or, as Davidson summarizes: “more ideas → better farming techniques (or other innovations) → more food → more people → more ideas → …” That feedback loop leads not just to economic growth but to accelerating economic growth.

This type of theory also explains why growth has slowed down in rich countries compared to where it was in the 19th century. In a process known as the “demographic transition,” people in richer countries tend to choose, for a variety of reasons, to have fewer children. This breaks the feedback loop because more ideas leading to more food no longer necessarily leads to more people.

But now, imagine that researchers are able to build two-legged robots, with hands and arms and everything, capable of performing both any physical task a human can and anything on a computer a human can. We’re talking full Blade Runner or Battlestar Galactica here (hopefully minus the rebellion).

We would be able to build these robots in a much shorter time than the decades it takes to birth, raise, and educate a human worker, and at less expense. So we’d achieve much faster population growth (or at least growth in the population of working robots) and bring back the feedback loop that caused economic growth to accelerate a few centuries ago. The fast-growing population of robots would be able to come up with, and implement, enough economically useful ideas to get the economy going faster and faster and faster.

The third argument for transformative growth is based on the conventional model that economists use to study growth in the medium to long run. The classic way of looking at economic growth is sometimes called the Solow-Swan model, after Robert Solow and Trevor Swan, who wrote separate papers developing it in 1956. (Solow died recently, in December 2023.)

In this model, the size of the economy — the amount of goods and services being produced in a given year — depends on the amount of labor, the amount of capital, and a measure of productivity. Capital here specifically means tools and property that can be used to make stuff: machines in factories, ovens and dishwashers at restaurants, trademarks and patents that represent ideas you can use to make stuff.

One of the most important aspects of this model is that there are diminishing returns to additional labor and additional capital. That’s because you need both to do anything useful. If you have a coffee shop with five baristas and no espresso machines, the first espresso machine is going to make them vastly more productive. But the 200th machine will do nothing because five baristas can’t run 200 machines simultaneously. Similarly, if you have 200 machines and no baristas, the first barista you hire is going to be enormously valuable. The 1,000th will be useless.

Put human-level AI into this model and a bunch of things can happen that make superexponential growth look likely. AI could, for instance, make returns to capital constant, rather than diminishing. That’s because you can always invest in capital (namely, robots or other AI) instead of labor and get the same effect as if you’d hired someone.

You can buy a robo-barista instead, and make all those espresso machines hum. That makes the labor component of growth literally irrelevant. Growth will explode. (Good.) But because demand for human labor will plunge to zero, most of humanity will be jobless and likely not share in that growth. (Bad.)

Economists Philip Trammell and Anton Korinek have reviewed some 25 ways of plugging AI into this standard model, as well as more recent “endogenous” models that treat technical change differently. Many of these approaches result in a prediction of superexponential growth. Advanced AI could automate research, fueling accelerating growth in productivity. It could increase the rate of return on investment in capital by making capital more useful (you have great robots now!), which spurs people to save more, which leads to more investment in capital, and so on. The exact mechanism varies based on the model and scenario, but it’s not hard to get the models to spit out a substantial acceleration in economic growth.

The models, of course, are just models, and inserting AI puts them “out of sample”: They’re designed for scenarios like the present, where human-level automation does not exist. But they’re also not just models: They express coherent stories and processes through which explosive growth could happen. It’s not hard to see how automating research, for instance, could lead to technology improving rapidly, with massive economic consequences.

“There is no shortage of mechanisms through which advances in automation could have transformative growth consequences,” Trammell and Korinek conclude, “once we allow ourselves to look for them.”

The basic reason to doubt a growth explosion

If the above all feels very speculative and theoretical, that’s fair. We have never had an AI-driven growth explosion before, and the effects of information technology on growth to date have been famously meager. In the US, the advent of personal computers coincided with a marked decline in productivity growth, not an increase. As Solow once put it, “You can see the computer age everywhere but in the productivity statistics.”

Beyond the surface-level sci-fi-ness of this narrative, though, economists and others have raised more specific doubts, many of which have less to do with what human-level AI would do than with whether we can achieve human-level AI any time soon.

In the above section, I asked you to imagine a robot in the style of Battlestar Galactica or Blade Runner, capable of doing all labor, both physical and intellectual, that a human can do. But we’re obviously a long, long, long way away from the existence of anything like that. Robotics has tended to lag behind software AI in recent years, and while some observers foresee that changing, it’s hardly guaranteed.

So it’s important to consider the economic impact of AI that can do most but not all of what a human can do. There are good reasons to doubt explosive growth in these scenarios, in particular because the scenarios strongly resemble what has happened in the US and other rich economies in recent decades.

One recent paper examined total factor productivity growth in the US between 1950 and 2018 and found that while it grew rapidly in some sectors (agriculture, durable goods manufacturing, wholesaling), it declined in others (construction, education and health care, finance/insurance).

This has decidedly not meant that the US economy has relied more and more on agriculture and manufacturing. In fact, employment in those sectors has fallen considerably, precisely because you can get more output per worker than in the past and so many fewer workers are needed to meet market demand. Automation has also led prices to fall in those sectors, and their share of overall economic output has fallen in turn.

By contrast, the share of jobs in those stagnant industries, the ones that aren’t getting more productive, has been increasing. And because the less productive industries are becoming a bigger and bigger share of the economy, overall productivity growth has been dragged down.

This is known as Baumol’s cost disease, after the late economist William Baumol, and it’s a dynamic that limits how much automation can supercharge growth. Even if you massively automate certain industries — and if you’ve been to a farm or car factory recently, you’ll have noticed that these facilities rely heavily on very sophisticated planters, combines, and industrial robots to automate many tasks — the same process will lead those industries to become a less important part of the economy, and the industries where progress is harder will become more important.

To apply this to the AI context, you can imagine AI leading to full or almost-full automation for a few tasks. Maybe it replaces front-end engineers for making websites and applications, or even software engineers en masse. Maybe it automates graphic design and 3D animation well enough that most businesses switch to using AI models rather than people. Maybe it replaces human journalists. (I’d prefer not, but I have my worries.)

As long as there are other jobs (chefs, child care providers, construction workers) where AI isn’t driving large increases in productivity — perhaps because we still can’t manage to produce useful robots that can put that AI into the physical world — the result of this process will not be explosive growth. The result will be that employment and prices in automated sectors collapse, those sectors become less important as a share of the overall economy, and economic growth as a whole is still bottlenecked by sectors where productivity growth is hard to achieve.

Jones, the Northwestern economist who has modeled how AI affects growth trajectories, anticipates that these kinds of bottlenecks will prevent explosive growth due to AI, at least in the near term. Think about how much technical progress has happened in computing over the past 70 years or so. “Moore’s law is almost absurd,” he noted in an interview. “It’s 10^17 more flops [a measure of computing performance] per dollar than 70 years ago. That’s incredible.”

But our ability to manipulate atoms hasn’t matched our ability to manipulate software bits, which is why, since the advent of the integrated circuit in 1958, economic growth in the US and other rich countries has not been explosive. There are other industries where productivity is not exploding, and those are holding us back.

“If you took a picture of a restaurant now and 1950, it’s effectively the same,” Jones says as an example. “They take your order, someone’s going to take an order to the kitchen, someone’s going to cook it using capital equipment and labor.” It might be a little cheaper now; the ovens and dishwashers are a little more efficient. But it’s not what we’ve seen with computers, and that’s meant overall growth has been modest.

Who’s right?

Believers in a growth explosion argue that modeling AI like this undersells its potential. Past technological advances, ones that have brought us steady but not accelerating growth over the past century or so, “took the form of technologies that automate small segments of production, offering modest benefits while requiring numerous expensive synchronized changes across the economy to be implemented,” economist and growth explosion theorist Tamay Besiroglu noted in a recent debate on the topic. “In contrast, if AI is capable of everything a human can do, we could potentially automate large numbers of tasks in one go, with fewer costly updates to existing processes.”

Notice here that Besiroglu is assuming an AI capable of everything a human can do. This isn’t strictly necessary for the growth explosion story. “It simplifies the argument to talk about full automation, but I think we could get explosive growth without literally full automation,” Davidson says. We don’t necessarily need to automate things like caregiving or teaching or surgery: “If you can fully automate R&D and capital investment, that gets the feedback loop going that gets growth going very fast.“

Fully automating research and development (R&D), of course, is no small thing either — and part of why this scenario sees fast growth is the R&D sectors are working hard to get around bottlenecks created by sectors that aren’t automated.

The more I dug into this debate, the more this seemed to be the crux of the disagreement. Believers in a growth explosion seem quite confident that it is possible, in a matter of decades, to develop AI and robots capable of doing any economically useful task a human can do, or any task important for the production of new ideas that drive productivity and economic growth.

Skeptics just don’t buy this. “This tech is amazing, it’s moving fast, it’s important,” David Autor, a professor of economics at MIT who has studied the effects of AI on jobs, told me. “But I don’t think it converges toward the end of labor.”

AI, as impressive as it is, is simply not on track to substitute for all labor, in this view. “AI does not reason,” Autor continues — which would, for instance, make it impossible to automate R&D. “It does not think analytically, it does not understand object constancy. I don’t think that problem solves itself.”

In some ways, this makes the question of whether AI will drive explosive growth a bit more tractable because there doesn’t seem to be as much disagreement among economists and other analysts about what human-level AI will do if we get it. The actual state of the technology seems like the biggest source of uncertainty, rather than the effects of its most extreme form. Human-level AI does seem very likely to drive explosive economic growth — by totally substituting for labor, by automating the discovery of ideas, or both.

If you think human-level AI is inevitable, this is both exciting and terrifying. Many of these explosive growth models project that demand for human labor will fall to zero. That’s a scenario of massive unemployment and grotesque inequality between the minority of people who own capital and profit from the growth explosion and the majority who lack capital and languish. Taxes and other mechanisms could seize some of the gains and redistribute them to the newly unemployed majority, but a scenario with unemployment levels far above those of the Great Depression would be rather ugly, alms or not.

Even if you don’t think human-level AI is possible or likely in the near term, the picture could still be interesting. There are many scenarios in which AI does not lead to an “explosion” in growth, or to superexponential growth, but does lead to growth being persistently higher for some time — and more widely spread. For instance, Autor is highly optimistic about the potential for AI to improve productivity in precisely those sectors (like health care and education) where it’s been stubbornly low, unclogging the bottlenecks that have been holding back the overall economy.

And because the unmet need in these areas is so high, he thinks this productivity could coexist with high levels of employment, unlike the situation in agriculture and manufacturing where high productivity has gone along with declines in employment. Health care is “not going to be like agriculture, where we have so much that it doesn’t employ anyone,” he says. “I don’t see it getting less labor intensive, but much more efficient.”

Explosive growth is a pretty high bar, even if its theorists make a compelling case that it’s at least possible. But even a smaller boost could wind up changing all our lives.

By .

Leave a Reply

Your email address will not be published.