How AI Might Impact the Economy—and What Government Could Do About It
Harvard economists map out scenarios, winners and losers, and possible responses.
According to a recent New York Times analysis, "the total value of the U.S. stock market has more than doubled over the past decade to over $75 trillion, roughly two and a half times the annual output of the entire U.S. economy, itself a record ratio." Much of that growth is driven by AI, or, more accurately, bets on future profits from companies developing and integrating AI—about half of the rise in the S&P 500 index this year was driven by AI-related stocks, the Times reported.
Given the stakes, Harvard economists Karen Dynan, PhD '92, Douglas Elmendorf, PhD '89, and Louise Sheiner, PhD '93, recently sought to outline some possible scenarios for AI's impact on the US economy, who the “winners” and “losers” would be, and, finally, how governments could mitigate negative effects through fiscal policy. Their analysis is distilled in "How Might Fiscal Policy Respond to the Rise of Artificial Intelligence?" a new working paper for the National Bureau of Economic Research. Dynan, professor of the practice in Harvard's Department of Economics and at the Harvard Kennedy School, recently sat down with Harvard Griffin GSAS Communications to discuss the paper, its recommendations, and her hopes for the work.
I’d love it if we could begin with the four scenarios for AI that you outline in the paper. What happens to economic growth, inequality, and jobs in each?
The first thing I should emphasize is that these aren’t predictions. Nobody knows what kind of economic outcomes are going to result from the rise of AI. What we’re doing in this paper is looking at four scenarios that are among the most often discussed ways in which things could play out.
The first scenario is the more optimistic, which is that AI raises productivity growth, which raises income growth, and that the income gains are broadly shared. In this case, the gains are distributed in proportion to where people started.
The second scenario assumes that the gains aren’t equally distributed—in particular, that the gains go to workers at the top of the wage distribution. So, you have faster economic growth from the increased productivity, but you also have a rise in income inequality.
The third scenario adds labor market disruption to rising income inequality. AI will cause some workers to lose their jobs, but it’s going to create new opportunities. Altogether, it’s going to create more churn in the labor market. This scenario assumes that it’s going to take some time for workers who’ve been displaced to find new jobs. So, you get some rise in the unemployment rate, and you get some workers dropping out of the labor force because they’re more discouraged as it takes longer to find a job.
The fourth scenario is the most disruptive. We have even higher productivity growth, but at the same time, we assume more displacement of workers, and in particular, we assume more permanent job loss. So, there’s a higher unemployment rate and more people dropping out of the labor force. The additional income in this scenario all goes to the owners of capital— the shareholders and other investors who own productive assets—not to workers.
The last two scenarios in particular remind me of the research your colleagues David Autor, PhD ’99, and Gordon Hansen did on the “China Shock.” Economists predicted there would be GDP growth, job growth, and new opportunities, and that the job losses relative to the size of the economy would be relatively small and manageable. What Autor and Hanson found was that those losses were deeply concentrated and had powerful effects on certain communities. Translated to our electoral system, the shock had ripple effects that people couldn’t really understand or predict.
The research around the China Shock and its effects on workers who lost jobs is important. The AI shock is unlikely to look exactly the same, as the China Shock resulted in highly localized economic devastation because it hit particular manufacturing communities very hard. Manufacturing is very geographically concentrated in our country.
AI is going to be a little bit different. It’s more likely to affect people spread across industries, occupations, and regions. Today it might be call-center workers and software coders, but tomorrow it could be administrative workers and lawyers. We may see substantial job loss, but it’s likely to be more diffuse geographically, even if it is larger on a national basis.
That said, the broader lesson from the China shock literature—and this is a theme that runs throughout our paper—is that job loss can be tremendously costly. In particular, we know that workers who lose their jobs because demand for their skills falls can face persistent earnings losses. The literature has documented effects on people’s health, the pain they are feeling, and family structure, as well as social and political dysfunction in the places where they live. We talk a lot about that in our paper because it is something that you want to consider in the policy response.
AI is going to be a little bit different [than the China shock]. It’s more likely to affect people spread across industries, occupations, and regions. Today it might be call center workers and software coders, but tomorrow it could be administrative workers and lawyers.
Technological advances usually hit low-skill workers. What’s different with AI is its ability to do work that has traditionally been thought of as high-education and high-skill. Is there any historical equivalent of that kind of impact with a new technology?
Not that I know of. A question that has been of tremendous interest to researchers is whether this will play out the same as the last big wave of technology, which benefited workers higher in the distribution but hurt those lower in the distribution. There’s been speculation that this time will be different. The literature is still very much in flux, but it suggests there could be a lot of heterogeneity—workers all over the income distribution and in many types of occupations will be affected. It’s hard to predict, which is precisely why we are looking at scenarios rather than just a single set of outcomes that reflects our view of the most likely thing to happen.
One real surprise for me reading your paper was the thought that all four scenarios you present would actually reduce the US national debt anywhere from 39 to 49 percent of GDP over 30 years. You say that not all of that will come from higher tax receipts due to economic growth. If not, what else is going on? This seems like terrific news, particularly at a time when the deficit and national debt are rapidly expanding. So, what’s the catch?
One important thing to recognize up front is that the numbers we calculated are based on the assumption that there is no change in fiscal policy in response to the rise of AI. Under current law, the faster economic growth associated with AI improves the budget situation. Debt remains high in all our scenarios, but it just doesn’t grow as much as currently projected by the Congressional Budget Office (CBO). Some of the improvement is because income is higher, meaning the government collects more taxes.
The CBO currently projects debt to rise to about 175 percent of GDP over the next three decades. Our estimates reflect how much lower that trajectory would be. In all our scenarios, debt remains above its current level of roughly 100 percent of GDP, but it rises by less than projected—ending up at roughly 135 percent in our first scenario and about 125 percent in the other three scenarios.
Another really important part of the story is where the improvement comes from. The higher tax revenues I mentioned earlier contribute a bit, but the bigger story is that, if fiscal policy is unchanged, many categories of government spending don’t automatically keep pace with a faster-growing economy. For example, Social Security benefits for people who are retired now are indexed to inflation, but GDP growth reflects productivity growth as well as inflation. So, the benefits don’t rise with the broader pace of the economy.
One catch is that history suggests policymakers don’t always allow those automatic changes to play out. Over time, they have often adjusted taxes and some categories of spending to keep them more closely aligned with the size of the economy.
People’s well-being depends not only on income but also on having a sense of purpose. If that remains true, you want to think about transition support, such as modernizing unemployment insurance, building better wage insurance programs, or improving worker training.
For the scenarios that have negative impacts—ranging from increased inequality to major disruption of the labor market—what can government do through fiscal policy to mitigate that impact?
We offer a menu of items rather than a particular package. Even in the optimistic scenario where the rising tide lifts all boats, we may decide we don’t want the government to shrink as a share of the economy—we might decide that our richer society deserves better roads or expanded access to education. If inequality grows, we will need to discuss whether we want a more progressive system of taxation.
Labor market displacement is another big area. Work is not just a way of bringing in income; it seems to have independent value to people. People’s well-being depends not only on income but also on having a sense of purpose. If that remains true, you want to think about transition support, such as modernizing unemployment insurance, building better wage insurance programs, or improving worker training. Finally, if income goes mostly to the owners of capital, the government could look at ways to own a piece of these companies—either directly or by setting up individual investment accounts.
Finally, if you were the chair of the White House Council of Economic Advisors, what would you urge leaders in Washington to do right now in anticipation of AI’s impact, and what is actually feasible in our political environment?
We want them to think about what types of policy would provide insurance against the more disruptive outcomes AI could bring. One example is strengthening and refining programs already in place to help workers transition to new jobs, such as modernizing the unemployment system and putting in place training programs backed by evidence. If bad outcomes occur, these programs will be there and can expand. Similarly, we could lay the groundwork for broader public ownership of capital that benefits from AI, so the public could share more directly in the gains if that became desirable.
As for the politics, preparing for uncertainty doesn’t require agreement about exactly what AI will do or how big a government you want. People with very different views can still agree that it’s prudent to improve our ability to help workers adapt if disruption occurs. I think that’s a much easier conversation than trying to tailor a set of policies based on predictions about what exactly is going to happen.
