AI and the Supply and Demand for Labor

I predicted a decline in net jobs due to AI, but not because of a jobs bloodbath

I was recently part of a panel of 16 economists asked to give forecasts about AI’s impact on the economy for the Wall Street Journal. What were the findings?

Every one of us agreed that AI would bolster productivity. Even folks who have been more skeptical publicly about the impact of AI agreed that it would have a positive effect.

But people disagreed about AI’s impact on labor markets.

5 economists said it would lead to net job loss, 8 said it would lead to no change, and 2 said it would lead to net growth.

I was one of the five, along with Daron Acemoglu, Rebecca Henderson, Pascual Restrepo, and Justin Wolfers, who said it would lead to net job loss.

These answers were highly correlated with other questions in the survey. For example, is AI more likely to replace workers or complement them? The same five of us said it would replace workers.

The five of us also agreed that AI would reduce demand for white-collar jobs.

So what’s going on here? Why do these economists give such different answers to these questions?

I think these questions actually are asking many subquestions that are getting rolled up in a way that leaves a lot open to interpretation.

Let’s focus on this question:

What informs how someone would answer this question?

I’m going to put these into buckets.

Definitional. What does net job losses or job growth across the economy mean?

• I interpret this to mean, will the prime age labor force participation rate go up or down because of AI?

• Perhaps you could have other interpretations. One very reasonable alternative is, on average will people work more or fewer hours in the future because of AI? Or, will AI lead to higher or lower wages?

But for now let’s take this to mean the prime age labor force participation rate. I predicted that AI would cause it to fall. But that immediately leads us to another bucket of questions.

The capabilities of humans versus AI:

1. Over what time horizon are we forecasting its impact on the labor force participation rate? Next 5 years, 10 years, 30 years, 50 years, ever? It’s unclear. I primarily considered a longer time horizon, say 50 years. Over that time horizon I expect it to reduce the labor force participation rate, all else equal.

2. How fast can AI capabilities improve? How quickly can we get autonomous robots? Again, it’s often confusing whether disagreements are questions about timelines or simply whether things are possible at all. I would be surprised if AI never advanced enough to perform most physical tasks, again all else equal. 50 years is a pretty conservative estimate for when people I trust expect highly capable humanoid robots.

3. What can humans do persistently (maybe in perpetuity) that AI can’t do? These get at the bottlenecks or weak links that maintain demand for human labor.

There can be plenty of discussion about the above issues. Some of the disagreement about AI’s effects on labor markets ultimately boils down to disagreements about the capabilities of humans versus AI, though people disagree less about this than you might expect! See the most recent results from the Forecasting Research Institute.

Given the improvement in and diffusion of the technology, any discussion about human and AI capabilities depends on the time horizon. In interpreting the question I considered longer horizons.

Other uncertainty:

Forecasting the economic impacts of AI requires making assumptions about everything else that happens in the future. This is a fraught exercise. Will the government take greater control over the development and release of AI technologies, which already seems to be happening? Will AI continue to be aligned with human interests? Will major non-AI social, technological, economic, or environmental changes occur? What about all the potential unknowable unknowns, changes to the world that can’t even be fathomed today?

My forecasts “hold all else fixed,” assuming the world continues along its current trajectory along every dimension as AI continues to develop. In reality, the current trajectory has zero probability of actually continuing in perpetuity.

I view the correct way to interpret these forecasts is as an instructive device. Construct the simplest model of the world that you believe captures the key levers for how AI could shape work. What does that model say about future labor force participation?

In truth, this isn’t really a prediction. I have a lot of uncertainty about the labor market impacts of AI. Even so, my uncertainty about labor markets is dwarfed by my uncertainty about the social, political, and other effects of AI.¹ I didn’t simulate all of these possibilities. I don’t think readers of the Wall Street Journal want me to give them a view about AI’s labor market impacts that averages over the probability that China invades Taiwan or that a future presidential administration bans data centers. I can certainly offer opinions, but I have more to say about the economy and the labor market than those possibilities. Instead, I just gave a summary of my model of the world for how AI might impact work, all else equal. All else will not be equal, and perhaps the difference between a superforecaster and an economist like myself is that their instinct is to try to simulate these alternatives, while my first instinct is to set them aside for the sake of instruction.

I can’t speak for everyone, but my impression is that most economists interpret projections in this way, as a simulation holding all else equal. It’s useful for tractability and for expressing simple but powerful ideas. But I would be hesitant to bet a bunch of money on these predictions because they rule out other changes that will occur with certainty.

Economic equilibrium responses to AI:

This is what I want to discuss the most, because it merits more attention than it gets.

Take a look at this back and forth. I think this bundles together a lot of key ideas that can get glossed over.

Both Andrey and Joshua are making good points here. Predicting “no change” literally is predicting something that has 0 probability of happening. There will presumably be some change in labor force participation caused by AI.

The people who predicted no net change probably took this question to mean “very small net change” or perhaps “very uncertain net change.”

Why might an economist have good reasons to predict “very small change”?

First ask the basic question. Why do people work?

People work because they prefer spending their time working compared to other things they could be doing instead.

There are two things to weigh when thinking about AI’s effect on work. How does the value of working change? And how does the value of doing something else change? Another way to phrase this is, how does the demand for labor change compared to the supply of labor?

Discussions about AI and work often muddle these things together in a way that is confusing.

Let’s start with labor demand, which gets more attention.

The effect of technology on labor demand balances two forces. On the one hand, it replaces tasks that are presently done by humans. There are countless examples of jobs that have shrunk because of automation, such as telephone operators, assembly line workers, or farmers. On the other hand, technology can increase demand for new or existing work, such as content moderators, rideshare drivers, or food preparers.

Empirically, the second effect has won out historically. Because humans perform tasks that are crucial to the production process and can’t be done by technology, demand for human work remains, as “weak links” or new tasks offset the effects of automation. Adding on rising productivity gives a credible explanation for why per capita incomes have been rising since the industrial revolution, even as tasks get automated.

So on net, labor demand and in turn wages have been rising, driven by productivity growth and weak links that offset the effect of automation.

Now consider the labor supply side, which often gets neglected.

Today, many prime-age adults have a strong incentive to work. If no one at home works, income and consequently household consumption would be far lower. This makes overall labor force participation pretty inelastic, at least at the household level.

Over the past 25 years overall labor force participation for prime age adults has been relatively flat. Prior to that it rose as women entered the workforce. For men it has been slowly declining, though it is still near 90%. The household labor force participation rate has been remarkably stable over time, even as rising female labor force participation had big effects on overall LFP. The share of households with at least one child under 18 that have at least one working adult has stayed around 90% for decades. Even as the US has gotten richer, people continue to work at high rates.

Data from FRED

Computed from CPS ASEC via IPUMS.

The above graphs suggest that labor force participation, the extensive margin of whether to work or not, has not notably declined overall since the 1960s.

This is one argument against AI leading to declining labor force participation. If people need to work in order to consume, then in equilibrium the adjustment won’t occur via whether people work. It will occur via either wage adjustments or hours adjustments.

Rather than ceasing work altogether people could just work fewer hours, as predicted by John Maynard Keynes and others. Perhaps the thing to focus on is hours worked rather than labor force participation. This is the intensive margin of how much to work. What does the data say?

There are notable differences in working hours across time and space, presenting a complex picture. Consider this graph of working hours per worker. The exact measurement details are hairy, but this should give a rough idea of how working hours vary.

From Our World in Data.

The graph suggests that people worked around 40% more in the mid-19th century than they do today across Western countries. In addition, today some places like the United States, and especially China, work more hours than in other countries.

What do economic models have to say to explain these patterns?

As societies get richer, there are two competing forces on labor supply: income and substitution effects.

Higher incomes make people richer. If people prefer to spend their time in leisure rather than work, then becoming richer should encourage people to tilt more of their time towards leisure and away from work. This is the income effect.

On the other hand, higher wages increase the payoff from working. If the next hour you work pays $10, you may or may not decide to take it. If the next hour pays you $1 million, you probably would. This is the substitution effect. Higher wages increase the marginal value of working.

The income and substitution effects go in opposite directions, so the net effect of rising incomes on labor supply is ambiguous. The graphs above suggest that higher incomes have led to people working less on average, roughly consistent with an income effect that outweighs the substitution effect.

The picture is of course more complicated, and varies a lot by context. Some of these changes in hours worked were driven by changes in regulation. The second half of the 20th century saw rapid increases in female labor force participation, which changes the composition of the workforce in ways that could affect average hours worked. In any case, people seem to work less today than they did in the past.

These averages mask important underlying shifts. In 18th and 19th century literature about European aristocrats, work is for the lower classes. The aristocracy attends balls and pursues hobbies like polo. Today, higher income workers actually work more. Why is that?

One likely explanation is that in the past the leisure class earned most of their income from capital, not labor. An additional hour of work might not have actually earned much money. Instead the real way to generate money was to own a bunch of land and collect rents on it.

This sort of capital income results in an income effect but not a substitution effect. People with high capital income from collecting rent from their land are rich, so they want to work less. Wages are also low, so there is little offsetting incentive to work more hours. In contrast, the lower classes had no capital income, so their marginal value of working was very high. This could explain why the poor worked more than the rich.

Today, the rich still benefit from higher capital income. However, the hourly wage rate for people at the top of the income distribution, especially those who work a lot of hours, is so high that the substitution effect outweighs the income effect. Richer people work more because they get paid a lot.

What changed? A likely explanation is that in the 20th century technological change was complementary to highly skilled, highly educated workers. As a consequence the college wage gap grew rapidly, encouraging educated, highly paid workers near the top of the distribution to work more.

Of course, these stories fail to resolve all of the differences in working hours in the graph above. For example, why is it that the US, the highest income country in the list, works more than all of the European countries, but works less than China, the lowest income country? Perhaps the people in different countries have different work and leisure preferences that lead to different income/substitution tradeoffs. Alternatively perhaps the labor market institutions and welfare systems differ. Probably all of these factors contribute.

A lesson is that labor supply decisions are influenced by a number of factors, leading to notable variation across time and space. On net, people seem to work less on average than they used to. This adjustment has come primarily via reductions in hours rather than labor force participation.

How might AI change things?

Now let’s put together supply and demand and think about the effects of AI.

Consider first the short run. Suppose that people need to work in order to consume, so that in equilibrium labor force participation would remain steady, with adjustments to hours and wages. In the short run there can still be acute impacts. Danny Yagan found that regions and individuals more exposed to the Great Recession faced longer-term and more adverse shocks. Other research finds similar results, from the replacement of routine work, the adoption of industrial robots, and from trade competition. Adjustments can be challenging, even if things eventually come back to equilibrium.

There is risk of “messy middle” scenarios where slow adjustments lead to disruption.² Switching to doing new things is hard. AI might make it easier or harder—at least to some extent this is a policy choice. Perhaps capabilities continue to advance rapidly, making it hard to find durable alternative work. Another more optimistic possibility is that AI creates new work, helps people learn new things, and eases transitions. Policy changes that reduce labor market frictions might help reduce impacts of automation. But there is risk of politically and socially volatile disruption that could even change long run trajectories. I don’t mean to minimize these possibilities, since they are perhaps even more important to think about.

But holding all else equal, how do we think about the longer run? Consider the following scenarios.

First, suppose that AI makes the economy much more productive but shifts production towards capital instead of labor. As a result, the government gives everyone a million-dollar UBI per year. In this scenario I bet fewer people would work. This would also be a huge boost to welfare, even with fewer net jobs.

A similar result can hold if instead of the government providing people with an income, people just own capital themselves and become wealthier over time.³ The capital owners might be the new leisure class.

Bear in mind that the outcomes are functions of policy choices. Consider another scenario. Suppose the government cares about encouraging work for its own sake. It might give a million dollars a year, but only to people who work 40 hours a week. Think of this as an extreme version of the Earned Income Tax Credit. In that case, a lot of people would work 40 hours a week, even if work looks very different from what we’re accustomed to.

This dependence on policy choices introduces uncertainty about future outcomes that hinge as much on the technology as the response to that technology.

Why did I predict declines in net jobs?

With this background, I can get into why I predicted declines in net jobs.

In the long run, I expect the economy to be much richer because of AI, all else equal. All 15 respondents predicted that AI would boost productivity.

I expect human input to be fundamental to the economy and society. If not, there are bigger things to worry about than the labor market! But many jobs that are done by humans today will not exist in the future.⁴

I also expect high returns to capital because of growth from AI.

Putting this all together, I expect labor force participation to decline in the long run.

I don’t think the likely driver is that it will be impossible to find work. I hope there are many people who spend their time ensuring AI and society at large are aligned with human well-being.

Rather, I expect in the long run work will be less fundamental to ensuring a reasonable level of consumption. I expect that AI will make goods cheaper and better as it automates swathes of the R&D and production processes. I also expect people will benefit from capital income that induces income effects but not substitution effects, either via directly owning capital themselves or by receiving payment from civic institutions or governments. I also expect that even if wages rise, eventually income effects win out against substitution effects.

Where could I be wrong? In lots of places.

Return to the statement I made above.

People work because they prefer spending their time working compared to other things they could be doing instead.

Why might people continue to prefer working? Some possibilities:

• They need to in order to consume. They don’t make enough money otherwise, either via saved labor income, capital income, or payments from government or civic institutions. One reason could be that AI doesn’t actually lead to a productivity take off because of persistent weak links, so the economy doesn’t actually become rich enough in the long run to support consumption without work for most households.

• They want to work for its own sake. The lines get blurred between work and leisure, so that most people just prefer working even if they don’t need to for basic consumption needs.

• Substitution effects grow even faster than income effects. Consumption becomes so good that people really want more money, even after their needs are covered.

A reasonable person could believe any of these things will be true in the future.

In any case, it’s worth emphasizing again that I don’t think it’s likely that labor force participation will decline in the long run because AI makes it impossible for most people to find jobs and drives them towards destitution.

If anything, I expect it to reduce labor force participation because it makes people rich enough that they don’t need to work in order to have a decent quality of life.

Some points I want to emphasize before concluding:

• I’ve made no statements about inequality. It could shoot up even if everyone gets richer. This influential piece from Philip Trammell and Dwarkesh Patel considers some scenarios.

• The short to medium run could still be challenging if transitions don’t go well. The long run might take too long to come, and in the meantime there could be disruptions and adverse outcomes. The marginal value of quality work exploring these possibilities is so high right now.

• Again, these are not predictions about the future as much as stylized instructive principles about how AI might affect work.

I hope this is a useful clarification about the many considerations that go into forecasts about the labor market effects of AI.

1. There are so many open questions even when just restricting to economic uncertainty. See my essay from October.

2. For example, my research with Erik Brynjolfsson and Ruyu Chen finds a slowdown in entry-level hiring in AI-exposed jobs.

3. Perhaps a hot take—if the marginal value of owning capital shoots up in a world with highly capable AI, more people will own capital, either directly or via their governments or other institutions. It should be a priority to make it easy for people to own stakes in the sorts of assets that will appreciate due to AI. Fintech innovations have already contributed to rapid growth in the number of unique Indian investors.

4. This is why I predicted a decline in labor demand for white-collar jobs.