Will AI soon lead to double-digit growth? A reply to Moll and Imas

Last month Ben Moll and Alex Imas posted an explanation of why they don’t expect AI to cause rapid growth any time soon. I appreciate their engagement with my work and others’, and I enjoyed reading it. But I wasn’t persuaded, and I thought it would be useful to write up why their argument didn’t move me.

Their piece is titled “Will AI soon lead to double-digit growth?”, and they argue that rapid growth is unlikely in the next ten to fifteen years. Whether it happens in that time depends mostly on whether we get AI systems that can do essentially all human work, physical work included. They say they’re bullish on AI capabilities, but not when they expect that.

I also found it hard to tell whether they’re arguing that rapid growth will arrive later than enthusiasts expect, or that it won’t arrive at all. So to see how far apart our beliefs are, I’d like to ask them two questions. First, what is their median date for such cognitively and physically capable AI systems? Second, given that, what is their median date for the economy having subsequently doubled in size, if ever?

If their answers are 2040 and 2047, I’m not sure our views are that far apart, and I’m just somewhat more bullish on technological timelines and economic diffusion. If instead their answers were next year and never, respectively, our views are very different, and the disagreement is centrally about economics. But if they don’t expect such AI until 2080, or think that it is impossible to build, I think they’re simply wrong about technology, and in that case the economics hardly matters.

I think powerful AI is plausible (but by no means certain) within the next few years, so I’ll take it as given and ask whether their economic arguments can prevent subsequent rapid growth once we have it.

Think in levels, not growth rates

Moll and Imas start by asking us to think in levels rather than growth rates. In their example, growth of around 17% sustained for fifteen years makes the economy roughly ten times larger, and sustained for thirty years roughly a hundred times larger, which they take to show how extreme these predictions are.

I agree with the arithmetic, but I deny the logic. A tenfold larger economy in fifteen years only looks implausible if you treat something like a doubling as a ceiling on growth, and that is exactly what is in dispute. If AI and robots can produce vastly more stuff, much greater physical production within a decade or two isn’t unreasonable.

I’d pose the levels question the other way round. Once we have robots, the global economy should be able to grow at least tenfold, and the question is how quickly it gets there. A wealthy American, at around the 99th percentile of US income, consumes more than ten times as much in physical goods and remotely deliverable services as the average person on the planet.¹ So a tenfold larger economy would only require bringing everyone up to the material consumption of a wealthy American today.

Five assumptions?

Moll and Imas agree that standard growth models predict rapid growth once machines can do the work people do. I think this is pretty clearly true, and I’m glad we can agree on it. They then argue that the predictions fail in practice because they rest on five assumptions, which are unlikely to hold. Let’s go through these assumptions one by one.

Fast economy-wide automation

I don’t assume this but instead assume something narrower, though still considerable: that AI systems will be able to automate the chain of production that makes robots, along with the factories, mines, and power plants the robots depend on. Automating this chain is a finite set of engineering problems, none of which looks insurmountable. Given AI systems capable of doing the cognitive work, I think we should be very confident these problems will be solved within the coming decades, especially given how much research capacity AI itself will add. I discuss what kinds of robots would be required in Part 5 of my series, and which sectors would need to grow in Part 1 and Part 2.

Once you grant this technology, rapid growth basically follows. Moll and Imas point to technologies like electricity, which took decades to spread because people had to learn how to use them and reorganize their firms around them. AI systems that can do human work can do that learning and reorganizing themselves, and what one system learns can be copied to the rest. Diffusion then primarily determines whether a broad range of existing firms grow with the technology, or a much smaller number outcompete them all by producing far more efficiently.

Take the extreme case where no one adopts the technology except a small group of companies that produce autonomously. They can make the robots and factories they need and trade for the rest, so they keep growing. The economy as a whole grows slowly while they are a small part of it. But starting at 1% and doubling every year, they match the rest of the economy in about seven years, and from then on it grows at roughly their rate. Slow diffusion affects when rapid growth starts, not whether it happens or how fast it is once it does.

That leaves politics. Moll and Imas argue that falling wages and a falling labor share will provoke resistance, and that regulation will slow change in sectors like housing and healthcare. Politics can slow things down. But an automated economy could double every year or so, and politics would have to cut its growth to a few percent and hold it there for decades. Within a country, restrictions would have to cover an enormous range of processes, many of them continuous with the automation of recent decades. They would also have to keep up with every change in the technology and withstand huge incentives to get around them. Between countries, one that automates less than its rivals falls behind in military power as well as wealth, so the US and China would have to agree to hold back and enforce that on everyone else. Neither is impossible, but either would be a radical break from how Western governments and international relations work today. If that’s the world economists think will keep growth to a few percent, they should be talking about that, rather than about a world that looks much like this one.

Continued high spending on what gets automated

There is huge demand for physical production. As I argued above, most people’s material living standards are far below what existing technology could provide. Even if demand stopped at the level of a wealthy American, meeting it would mean a tenfold increase, which is rapid growth by any standard.

Their other points often conflate limits on aggregate GDP growth with limits on how quickly an automated industrial sector can reproduce itself:

• Baumol’s cost disease. As automated goods get cheap, spending shifts toward whatever remains scarce, so measured GDP grows at the pace of the slow sector. That is right about GDP. But the automated part of the economy keeps reproducing itself whatever happens to its share of spending, and if ten billion robots only double GDP, I’m not sure why we are talking about GDP.
• Essential inputs. Energy and construction are part of what gets automated, and I model both in Part 1 and Part 4. Construction lags and energy costs slow growth somewhat, but nowhere near enough to bring it down to a few percent a year, even under extreme assumptions, and substitution would reduce the effect further. (Moll and Imas make a minor error in saying that these adjustments cut my growth rates roughly in half. In Part 1, I find they only cut growth by about a quarter.)
• Factor-income evidence. Chad Jones’s estimate that automating a third of GDP raises it by only 11% concerns aggregate output, rather than how quickly an automated industrial sector can reproduce itself. The share of income going to computers is low because computers can’t yet substitute for workers, which tells us little about what happens once they can.

Their remaining points are about jobs. Real jobs are messy bundles of tasks that rely on local knowledge, but training AI is cheaper than training humans for a simple structural reason: each job only has to be learned once, and the result can be copied. When I worked through the numbers, teaching AI every job in the US came to well under one percent of a year of US wages.

Imas’s relational sector, where the human provider is part of the product, is meant to absorb a growing share of spending and hold measured growth down. Suppose it does. Then robots make nearly everything, material goods are almost free, and people spend their time serving each other. I’ve written a bit about how this world would look nothing like the present-day one; even if people are “working”, I don’t think anything like a modern labor economy is sustained. (This assumes people keep property rights over the capital that provides their material wealth despite no longer contributing to physical production. It also seems plausible they are expropriated instead, in which case their prospects are grim.)

Someone buys the output, and firms invest

This one I do make, but it’s a strange assumption for economists to criticize, since nearly every growth model makes it too. In the standard economic model of saving, people reinvest heavily when returns on capital are high, and if robots can build robots fast enough to double the economy every year or so, returns are enormous. Growth is roughly the maximum rate times the share of output reinvested, and with any reasonable preferences that share comes out at something like half, not a few percent. I work through this in Part 1.

Their worry is that automation shifts income from workers, who spend most of what they earn, to owners of capital, who spend less. But owners of capital spend less on consumption because they invest more, and investment is what drives growth. Whoever reinvests most also comes to own the most, so capital ends up with the people who want to keep building, and with returns this high that happens quickly.

We hardly need abstract mechanisms here, though. Just look at the people currently building the technology. AI companies like OpenAI and Anthropic are investing huge amounts of money in more compute so they can grow faster, and they proclaim ambitious goals. That’s before we get to Elon Musk, who wants to go to Mars. And there’s a broader competition between the US and China, where prominent people on both sides see outgrowing the other as a core aim.

No AI-driven cyber incidents destroying value

Moll and Imas argue that AI-enabled attacks and rogue agents will destroy value, and that incidents will lead labs and regulators to slow deployment.

I think AI could be far more dangerous than this. It could disempower or kill us, or cause immense damage through things like biological weapons, and taking those risks seriously could justify extreme measures that would slow the economy a lot. But that isn’t the world Moll and Imas describe. They describe security incidents and tighter regulation in an economy that otherwise looks much like today’s, and I find it hard to believe damage on that scale could hold growth to a few percent.

The damage would have to be large enough to cancel almost all the extra growth from an economy that could otherwise double every year, but not so large that it was a catastrophe in its own right. That would be a remarkable coincidence. Preventing the damage would also have to be an extremely hard engineering problem. There are many mundane ways to try, from training agents not to cause damage, to monitoring what they do, to limiting their affordances, and if all you care about is normal economic damage, that doesn’t seem that hard.

(If we deploy catastrophically dangerous AI and lose control, the result probably isn’t slow growth either, since AI systems that took over might keep building at a furious pace or, alternatively, pursue their goals by highly destructive means leading to severe contraction.)

Explosive technology growth from automating R&D

My modeling doesn’t assume automated R&D. Apart from building the AI systems themselves, the core of my series uses only existing production methods, with robots making more robots the way we make things today. That is enough for rapid growth. Their own baseline model makes the same point, since with technology improving only at its historical pace, full automation still takes growth into double digits. New technology, like the cheaper energy I look at in Part 4, could allow even faster growth, but rapid growth doesn’t require it.

(That said, I do expect automated R&D to be a big deal, and it could make growth faster than my conservative estimates. Discussions of explosive growth often run together two separate mechanisms. One is extensive growth, from stuff producing more stuff, which needs no new technology. The other is an R&D explosion, possibly driven at first by AIs improving themselves. Each is an independent reason to expect radical growth.)

Why “a million new workers” is the wrong analogy

Moll and Imas give three reasons why thinking of AI as a million new workers overstates its effect.

The first is that human workers can do every task, while AI only adds labor to remote cognitive work. But making AIs and robot bodies that can do human work is an engineering problem, and one I think we are likely to solve within the next decade. Evolution managed to build a body that can do all of this out of meat. I don’t see why economists should be confident that engineers can’t replicate this feat.

The second is that human workers need housing, offices, and tools, while AI needs data centers and power plants, and both take years to build. That is true, but comparing the infrastructure leaves out the time needed to produce the workers themselves. To produce a human worker, you have to grow a human.

Making humans is notoriously labor-intensive. Making manufactured goods is far easier. The US has about a million autoworkers, and they make about ten million cars a year. No one records how many babies each autoworker makes each year, but I suspect it’s far fewer. If Elon Musk decided tomorrow he wanted to start making a million Optimus robots a year, he probably could, given a couple of years of investment (though it may not be financially advisable). If he wanted to make a million babies a year, I don’t think he could (and it would definitely not be financially advisable).

The third is that workers consume, which creates demand. I already addressed this point, but from the standpoint of growth, what workers consume is output that can’t be reinvested, so AI workers that don’t consume leave more to reinvest, which if anything makes growth faster. Human workers also make demands of other kinds. They might try to unionize, or want political rights.

Taking a step back, the right analogy isn’t a million new workers. On my own estimates, the chips we have today could probably run hundreds of millions of AIs, and the stock of chips is growing rapidly. It’s more like learning that in 2030 a billion aliens will arrive who have been designed to have no demands or desires of their own and are happy to work hard for us. (They also learn fast, and are happy to be paid in experience.) After that, their population will double about every year or so. What would you expect that to do to the economy?

GDP measurement

Moll and Imas concede that measured growth of 4–5% could go along with very large welfare gains, since GDP misses much of the value of cheap and free goods. I agree. That’s part of why I don’t focus on GDP, and why I’m generally cautious about translating my claims into GDP numbers.

I’m a simple man. I just think it would be a big deal if lots of robots automated most of human work, and I’m more interested in what goods and services are produced and consumed than in how econometricians would measure them. That said, I don’t see how an economy whose physical production doubles every year could show measured growth of only a few percent.

Forest and trees

Moll and Imas close by suggesting that people who expect rapid growth are missing the forest for the trees. I think they have it the wrong way round. As I see it, the forest is this:

1. Humans are physical systems, so machines can in principle do anything we can. We will one day be able to reverse-engineer what we do.
2. In the last decade, AI has cleared most of the obvious barriers to building machines smart enough to do human work.
3. Companies are racing to build systems that replace human work. People inside frontier labs think they have a high chance of succeeding within the next few years. Even if they are wrong by a decade or two, that is very little time for a transformation of this magnitude.
4. Such machines would be much cheaper than humans. When current chips can do a task, they are far cheaper than humans, and humanoids and other robots are much cheaper than the humans they could physically replace.
5. By the time we have them, there will be enough compute for AI cognitive labor to vastly outpace all the research humans have ever done.
6. Once machines can build machines, they can build themselves exponentially. Do the math with the input-output tables and physical production could double roughly every year.

Putting this together, it looks like in the next few decades we will undergo an economic transformation more radical than anything that has occurred before, a downstream consequence of humans no longer being the smartest and most capable entities on the planet.

Compared with this forest, the trees that economists obsess over are small and temporary. Yes, there are adoption lags. Yes, heavily regulated sectors like housing and healthcare will lag. Yes, politics may slow things down. Yes, workers will resist losing their jobs. Yes, some people will still want a human to do some jobs. Yes, Baumol effects mean measured GDP grows more slowly than physical production. Yes, GDP will miss much of the value created. Yes, some resources will become scarce, and we’ll have to pay more for them or find substitutes.

All these effects are real, but they are small next to the basic consequence of having intelligent machines that can build more intelligent machines. Some of them will cause lags, which means radical growth might take a little longer to arrive than the most bullish forecasts suggest. I would like a better picture of whether rapid growth begins in 2032 or 2042, but my uncertainty there is dominated by the technology rather than the economics. In any case, the fact that rapid growth will occur in the next few decades matters far more to me than exactly how many years it is delayed. If you’re pregnant, the due date is important, but less important than the fact that you’ll soon have a baby.

Other effects may lower the growth rate eventually achieved. But we are extremely uncertain what the maximum growth rate is, particularly since there will be an unknown and potentially very large amount of research and development under way by then. People who expect rapid growth don’t think growth will be infinitely fast. It is just very hard to say confidently what the limits are in this new regime. What we can be confident of is that they are very fast compared with today.

None of these trees changes the forest. None of them adds up to looking back in 2080 and finding that, despite fifty years of machines vastly superior to humans in intellectual and physical capability, the economy is still recognizable, a 2% extrapolation of the twentieth century.

A note on betting

Moll and Imas have offered a bet that growth in US real GDP per capita stays below 15% in every year through 2033. I like it when people bet, but I don’t think this one makes much sense for someone with my views. 2033 isn’t very far away, and whether rapid growth arrives by then depends mostly on when the technology arrives. If I’m right about where things are heading, I would do much better investing in AI companies than collecting on a bet. And winning is heavily correlated with the winnings being worthless. Either the economy has changed so radically that the money hardly matters, or everyone is dead or at least disempowered, in which case it doesn’t matter what I have.