A Bullish View of AI That’s Not So Bullish

Most of us see the rapid increase in the capabilities of generative AI as displacing knowledge workers, rendering most of them economically superfluous and, thus, destitute. An Economist feature asks, “What if AI made the world’s economic growth explode?“

The setup:

UNTIL 1700 the world economy did not really grow—it just stagnated. Over the previous 17 centuries global output had expanded by 0.1% a year on average, a rate at which it takes nearly a millennium for production to double. Then spinning jennies started whirring and steam engines began to puff. Global growth quintupled to 0.5% a year between 1700 and 1820. By the end of the 19th century it had reached 1.9%. In the 20th century it averaged 2.8%, a rate at which production doubles every 25 years. Growth has not just become the norm; it has accelerated.

If the evangelists of Silicon Valley are to be believed, this bang is about to get bigger. They maintain that artificial general intelligence (AGI), capable of outperforming most people at most desk jobs, will soon lift annual GDP growth to 20-30% a year, or more. That may sound preposterous, but for most of human history, they point out, so was the idea that the economy would grow at all.

The likelihood that AI may soon make lots of workers redundant is well known. What is much less discussed is the hope that AI can set the world on a path of explosive growth. That would have profound consequences. Markets not just for labour, but also for goods, services and financial assets would be upended. Economists have been trying to think through how AGI could reshape the world. The picture that is emerging is perhaps counterintuitive and certainly mind-boggling.

My initial instinct is that, while this would be great for “the economy,” it doesn’t really help tens of millions of people who are displaced from relatively pleasant, well-paying jobs. But, of course, new technologies have been displacing occupations for centuries.

Economies originally grew largely through the accumulation of people. Bigger harvests allowed more mouths to be fed; more farmers allowed for bigger harvests. But this form of growth did not raise living standards. Worse, famine was a constant menace. Thomas Malthus, an 18th-century economist, reasoned that population growth would inevitably outstrip agricultural yields, causing poverty. In fact, the reverse occurred: more people did not just eat more, but had more ideas, as well. Those ideas led both to higher output and, eventually, to lower fertility, which set output per person climbing. AGI, the theory runs, would allow for runaway innovation without any increase in population, supercharging growth in GDP per person.

Most economists agree that AI has the potential to raise productivity and thus boost GDP growth. The burning question is, how much? Some predict only a marginal change. Daron Acemoglu of the Massachusetts Institute of Technology, for instance, estimates that AI will lift global GDP by no more than 1-2% in total over a decade. But this conclusion hinges on an assumption that only about 5% of tasks can be performed more cheaply by AI than by workers. That assumption, in turn, rests in part on research conducted in 2023, when AI was less capable.

More radical projections of AI’s economic impact assume that much more of the world’s economic output will eventually be automated as the technology improves and AGI is attained. Automating production then requires only sufficient energy and infrastructure—things that more investment can produce. Usually, investment-led growth is thought to hit diminishing returns. If you add machines but not workers, capital lies idle. But if machines get sufficiently good at replacing people, the only constraint on the accumulation of capital is capital itself. And adding AI power is much faster than waiting for the population to expand, argues Anson Ho of Epoch AI, a think-tank.

This will make a lot of people very rich. But, again, it’s not obvious how the rest of us benefit.

This, too, seems less than exciting from a knowledge worker perspective:

Truly explosive growth requires AI to substitute for labour in the hardest task of all: making technology better. Will it be AI that delivers breakthroughs in biotechnology, green energy—and AI itself? AGI agents will, it is hoped, be able to execute complex, long-running tasks while interacting with computer interfaces. They will not just answer questions, but run projects. The AI Futures Project, a research group, forecasts that by the end of 2027, almost fully automated AI labs will be conducting scientific research. Sam Altman, the boss of OpenAI, has predicted that AI systems will probably start producing “novel insights” next year.

Economists who study “endogenous” growth theory, which attempts to model the progress of technology, have long posited that if ideas beget more ideas with sufficient velocity, growth should increase without limit. Capital does not just accumulate; it becomes more useful. Progress is multiplicative. Humans have never crossed this threshold. In fact, some economists have suggested that ideas have become harder, not easier, to find over time. Human researchers must, for instance, master ever more material to reach the frontier of knowledge.

Again, this sounds like it would be great for society, but bad for those who make a living doing science.

What would all this mean for workers? Humanity’s first growth surge was not especially generous to them. An English construction worker in 1800 earned the same real wages as one in 1230, according to Greg Clark of the University of Southern Denmark. The growing number of mouths to feed in effect nullified all the increase in output. Some historians argue that over the following 50 years or so, workers’ living standards outright declined.

What’s 50 years, right? A Keynes said, in the long run we’re all dead.

This time the worry is that workers become redundant. The price of running an AGI would place an upper bound on wages, since nobody would employ a worker if an AI could do the job for less. The bound would fall over time as technology improved. Assuming AI becomes sufficiently cheap and capable, people’s only source of remuneration will be as rentiers—owners of capital. Mr Nordhaus and others have shown how, when labour and capital become sufficiently substitutable and capital accumulates, all income eventually accrues to the owners of capital. Hence the belief in Silicon Valley: you had better be rich when the explosion occurs.

So, again, this isn’t making me feel better.

A booming but workerless economy may be humanity’s ultimate destination. But, argues Tyler Cowen of George Mason University, an economist who is largely bullish about AI, change will be slower than the underlying technology permits. “There’s a lot of factors of production…the stronger the AI is, the more the weaknesses of the other factors bind you,” he says. “It could be energy; it could be human stupidity; it could be regulation; it could be data constraints; it could just be institutional sluggishness.” Another possibility is that even a superintelligence would run out of ideas. “AI may resolve a problem with the fishermen, but it wouldn’t change what is in the pond,” wrote Philippe Aghion of LSE and others in a working paper in 2017.

So, maybe the boom will hold off long enough for me to retire. Great! But not so great for my kids.

Indeed:

Averages conceal variation. Explosive wages for superstars would not console those with more mundane desk-jobs, who would have to fall back on the parts of the economy that had not been animated. Suppose, despite AGI, that technological progress in robotics were halting. There would then be plenty of physical work requiring humans, from plumbing to coaching sports. These bits of the economy, like today’s labour-intensive industries, would probably be affected by “Baumol’s cost disease” (a wonderful affliction for workers) in which wages would grow despite a lack of gains in productivity.

In the classic case, named after an economist called William Baumol, wages grow to stop workers switching to industries in which productivity is surging. That would not apply with AGI, but other factors might produce Baumol-like effects. AI-owners and elite workers might spend a good deal of their new fortunes on labour-intensive services, for example. Think of today’s wealthy, who shell out on lots of things that are hard to automate, from meals in restaurants to nannies. It is an optimistic vision: even those who are not superstars still benefit.

This is great if your ambition is to prepare meals and provide childcare for the rich. Otherwise, not so much.

The non-rich would enjoy only selective abundance, however. Their purchasing power over anything that AI could produce or improve would soar. Manufactured goods made in AI-run factories could be close to free; riveting digital entertainment might cost almost nothing; food prices, if AI worked out how to increase agricultural yields, could collapse. But the price of anything still labour-intensive—child care, say, or eating out—would need to rise in line with wages. Anyone who switched from today’s knowledge work to a labour-intensive alternative might find that they could afford less of those bottle-necked goods and services than they can today.

You think?

There’s more to the piece, but it doesn’t get more hopeful.