Bill Kerr: Will artificial intelligence level the occupational playing field or increase inequality? Early studies in generative AI have pointed to greater gains for less expert workers. But subsequent research suggests top performers may widen their advantage. What can we learn from previous technology revolutions?

Welcome to the Managing the Future of Work podcast from Harvard Business School. I’m your host, Bill Kerr. My guest today is economist David Deming, professor at the Harvard Kennedy School and the Harvard Graduate School of Education. David is the faculty co-director of Harvard’s Project on Workforce and a principal investigator in the CLIMB Initiative on higher education and social mobility. In a late 2024 paper, David and co-authors, Larry Summers and Christopher Ong, looked at the history of occupational churn and what it reveals about current trends, including the influence of AI. We’ll talk about the implications for jobs, education, training, and public policy. David first joined us in 2023 to discuss college-to-workforce transition. David, welcome back to the podcast.

David Deming: It’s great to be back with you, Bill. Thanks for having me.

Kerr: David, we covered a lot of your background on the previous podcast, but maybe give us a little bit of a snapshot, including the CLIMB Initiative, which I don’t think we spent any time on earlier.

Deming: Yeah, sure. So as you said, I’m a professor at the Kennedy School and the Graduate School of Education. I’m an economist, and I study labor markets, education, and the intersection of those—so skills, technology, the future of work, training technology, and increasingly AI. And the CLIMB Initiative, just to answer your question, is an initiative that I started several years ago now with Raj Chetty and John Friedman, which seeks to understand the role of colleges in making economic inequality better or worse and how they can improve on some metrics related to social mobility. So we released a paper studying admissions at elite colleges. And we’re working on a project that is ongoing, looking at the broader landscape of colleges. So what are colleges doing to promote social mobility, which colleges are doing better and worse, and can we connect that to certain practices or policies? And workforce is likely to be a big deal with those colleges as well.

Kerr: Absolutely. In the first podcast, we talked about college-to-workforce transitions. A couple of years on, how is that looking, any different from where we were in 2023 in terms of placement patterns?

Deming: Yeah, it’s a great question, Bill. What we actually see in the data really depends a lot on what types of colleges you’re looking at. So if you look at community colleges, they’ve actually been the recipient of a lot of federal funding and a lot of support through a variety of policies, and yet, enrollment in community colleges declined a lot during the pandemic and is much lower than it was a decade ago. And four-year college enrollment is actually very slowly ticking up. The colleges where the admissions rates are still very low and lots of people still want to go. And it’s actually the colleges that people like—most people give very high approval ratings to their local community colleges—it’s those colleges that are actually suffering enrollment losses.

Kerr: What are the leading rationales or hypotheses even toward the community college downturn? Is it like a strong labor market and lower unemployment, so that’s boosting wage opportunities people can get right away?

Deming: Yeah, that’s a big part of it, Bill. You’re exactly right. If you look at historical patterns of community college enrollment, it is highly countercyclical, meaning when the economy is bad, people go back to college, and when the economy’s good, they tend to stay in their jobs. So part of it’s that. It’s not all that. I think what’s happening with community colleges is that they’ve always, for a century almost, had a dual mission of the transfer function, which is helping people go to the first two years of a four-year college at lower cost and then transferring, and then the workforce mission. Every college is trying to do more workforce things, but it’s really a matter of prioritization and investments. And so what I suspect you’re seeing is people increasingly doubting that the community college enrollment is worth the opportunity cost of their time unless they can clearly link it to a job downstream.

Kerr: You also talked about the structural changes. I think I read something recently, it said this year was going to be the peak year in terms of domestic enrollment in the colleges and it’s expected to decline, although potentially foreign students could keep the levels the same or higher going forward.

Deming: Yeah, that’s right. And so it’s the community college and the less selective directional schools that are going to face more competition to fill seats. Colleges are notorious for not being very good at lowering costs, and in their defense, a lot of their costs are fixed costs—salaries, tenured professors. And so that’s a real challenge for a lot of schools. You have a pretty fixed cost base, and you’ve got declining revenue, because you’ve got fewer students coming in. So I think you can expect to see some college closures and some restructuring coming up in the next decade or so.

Kerr: Well, around Harvard Business School, at least, the one way we’re most thinking we can lower costs is by trying to harness some of these technological capabilities in the workplace. So you had this work that you did with Larry Summers and Christopher Ong and just tell us about the origin of this paper. It focuses on both a long horizon of technological change and occupational change, but then also kind of narrows in on GenAI and some thoughts about it.

Deming: So the origin of this paper comes from a conversation that Larry and I have been having, actually for almost a decade now. We really started talking about this back in 2016 or 2017 when—I don’t know if you remember then—but there was a lot of conversations about technological unemployment. There was this paper written by Carl Benedikt Frey and Martin Osborne, where they claimed that up to half of all jobs in the U.S. could be automated within a couple of decades. There’s a lot of automation anxiety in the public. And as economists, we tend to be skeptical of the “This time is different” arguments.

Kerr: You tend to be skeptical of just about everything.

Deming: Yeah, that’s fair, fair, fair enough. But I think in this case, it’s warranted, right? Because if you look at the history, there are articles. Lyndon Johnson in 1965 had a blue-ribbon commission on automation in the workforce, and jobs didn’t disappear. Going all the way back to the Luddites and the looms, there are lots of examples in history of people thinking that technology was going to take all of our jobs, and it ends up taking some jobs, but then creating new jobs, and on net, there’s more people employed than ever. So back in 2017, we constructed this measure we call “occupational churn” that you alluded to in the intro. The idea is really very simple. We take all the jobs in the U.S. economy at any one period of time—let’s say 2020 or 2010—and we compute shares for each job type, for each occupation. So, managers, farmers, production workers. So you create categories for all the occupations, and each one has a share associated with it, and the shares sum up to one. So managers are 15 percent of the workforce and farmers are 5 percent of the workforce, and so the shares sum up to one, meaning you can compare the percentages over time. Think about it as if I randomly picked a worker in the U.S. economy, what is the probability they’d have each one of these jobs? So going all the way back to 1880, we have data on this, and the census is administered in the U.S. every decade. So you’ve got 1880, 1890, 1900, 1910, all the way up to the present. You can compute these occupation shares. And there’s a little bit of an art to it, because the job of software developer or occupational therapist didn’t exist in 1880. So you have to create slightly broad categories to daisy chain these things together.

Kerr: Yeah, I was going to specifically ask you this kind of nerdy question. When you have new occupations come in completely, do you capture that as just something brand new, or are the same kind of 10 buckets being used constantly since 1880 coming forward? Instead, it’s information processing and not something specific about software?

Deming: So a job like software developer would be a zero in all of the years from 1880 to 1970, roughly, and then it would start to have positive mass. Let me just tell you about what the measure of churn is. So you’ve got these probabilities in each decade, and then what you do is, between one period of the next, you want to know how different is the structure of the job market, how much has it changed? And one way to measure that is to say, “Well, if managers,” let’s say, “were 10 percent of the economy in 1980 and 12 percent in 1990, they’ve gained two percentage points.” And by the way, the loss has to come from somewhere else, because they always sum up to one. And so, what we do is, we just take the difference between, in that case, the 12 and the 10, and we take the absolute value of that. So if it went from 10 to 8, so if managers declined, we’d also call that two percentage points of change. So you sum those changes up over all the occupation categories, including the zeros, and that’s your measure of churn. And the virtue of that is that, because it sums to one in every year, it’s on the same scale, and what it says is if the job market’s totally stable, you’re going to have a measure of zero, so exactly the same proportions, and the more churn between categories there is, the bigger the number gets. And so again, it’s not a perfect measure of job change. There’s things like moving from the farm to the cities that are disruptive but aren’t captured in the same way and so on. But the virtue of it is it allows us to basically compare labor market disruption back 150 years almost.

Kerr: I’m going to hold on the punchline of where that trend’s going for just a little bit longer. And so how do we think about the tasks that could sit within an occupation change happening? So you’re going to capture the broadest movements, but to the question that people have—"Is this time different?”—do we have any reason to think that that measure would look different if we had it at a task level?

Deming: So if you have the job of manager or doctor, it implies a certain set of tasks you do, and so if you switch from one to the other, that is a change in tasks in the economy. Having said that, I think the question behind your question is, isn’t there a lot of changes within occupation and tasks? I would love to measure that. We just can’t really do it with the data we have. My suspicion is that that would increase the overall sense of churn in all years. There’s more disruption. I don’t know how it would affect the relative comparisons across years. I think that’s hard to say. That’s a great question.

Kerr: All right. So we’re down to that billion dollar question, which is, “Is this time different?” And you actually frame it even just a little bit, I think, narrower in the initial one, which is the change getting faster? So there’s both different, there’s also is it even getting faster than it was before?

Deming: Yeah, okay. So here’s the punchline. So again, as I said, Larry and I dreamed up this measure in 2017, roughly, and we actually calculated it in 2017, I think it was 2018. It was 2010 to 2018, an eight-year period. What has the churn been like, and how has that compare to the past? And what we found is that 2010s were the most stable period in the history of the U.S. labor market. And if you go and look at the past, the most unstable period, the period where you had the most disruption, was the middle of the 20th century. What was happening then? It was the post-war period, we had a huge manufacturing boom. The personal automobile disrupted the railroads. So a lot, like my grandfather, worked on the railroads, and those jobs went away. Railroads became mostly commercial, and you had a bunch of work in auto mechanics and auto repair. And then you had the rise of the office, so people were moving off the farm and into the office. The mechanization of farming really took off in the 1940s and 1950s. People finally gave up their livestock and bought a bunch of tractors. The job share of farming went down a lot. So all these things were happening in the post-war period, and that disruption is just much greater than what we’re seeing today in this sense of term that we calculated. When we did that, again, back in 2017 or so, we kind of patted ourselves on the back, we were smug about it, “No, this time isn’t different. The cranky skeptical economist wins again.” We kind of put it on the shelf, we never wrote it up. We just congratulated ourselves on this finding. So then we had the opportunity to write this paper over the summer that we’re talking about now. It’s 2024, so we say, “Okay, let’s rehash this, and let’s do it seven years later, let’s update the data, and let’s write the paper and see what happens.” It turns out that when you recalculate these measures all the way up to 2024, incorporating the post-Covid labor market, things actually look like they are changing faster. So the punchline at the end of the paper is, if you see the same measure of churn from 2010 to 2024, you see that the labor market is changing faster than any period since the 1970s. And so the paper is really two things. One, it’s, “Hey, in the broad perspective, everybody calm down.” And the second part is like, “But if you wanted to tell a story that this time might be different, you might see some early signs that things are changing more rapidly.”

Kerr: Yeah. So now you’ve outed yourself as both that skeptical economist but also the two-handed economist.

Deming: I’m the two-handed economist, yeah.

Kerr: Let me start with the long-run perspective. I guess I have two questions there. One is, I want you to tell us a bit about general-purpose technologies, because that’s going to set up GenAI. And the second is, just, it strikes me that one of the distinguishing features of that era was so many people, more than 50 percent, being in agriculture and coming out of one particular sector. And I guess, unless we all become minions to some robot overlords, we’re probably never going to again have 50 percent workforce in one category again. And so is there something that we should discount about that earlier episode? I’m just curious as to how you think about that, such a seismic change, compared to now everything is much more spread out in general.

Deming: Yeah, it’s a fair question, Bill. So this is where this whole thing becomes more of an art than a science. Why was the change out of farming so disruptive? Well, it was because a lot of the farming was subsistence farming. So it wasn’t farming to sell your crops on the market, it was actually just producing enough to feed your family. And that was kind of the state of nature for most of human history. And so the Industrial Revolution was really the ability to specialize in a certain kind of thing that by itself didn’t provide people with sustenance, but allowed you to trade your skills in exchange for other things that would sustain you and your family. And it’s probably right, that that’s a much bigger change than anything we are likely to see now, and so maybe it’s not a totally fair comparison. I think that’s a fair point.

Kerr: And then how do we think about the general-purpose technologies? Do you mark most of these waves as being connected to general purpose technologies?

Deming: Yeah, so that’s the other interesting thing about looking at this in the very long run perspective, Bill, is that, if you look at occupational change or jobs—people in jobs doing worse or better over a five-year period or a 10-year period—it’s really hard to know how much of it is technology, versus changes in institutions, the minimum wage, unions, sectoral bargaining. There’s a lot of other things happening, and so you can argue about what is the real source of changes in jobs, changes in work. But over the very long run, it’s just very clear that the primary thing is technology. So it becomes very important—if you’re thinking about “Is this time different?”—to ask the question of whether AI falls into the kind of “most disruptive” category of technologies that economists have, which is what you say, the general-purpose technology. What are some other general-purpose technologies? Electricity is a big one. Steam power was another one. The main thing is, it’s general purpose, meaning it applies not just in one sector. So you might have a technology like a typewriter. It’s a great technology, it unlocked productivity gains in office work, but the typewriter didn’t help farmers be more productive. It didn’t necessarily help musicians play more music or whatever. It was kind of limited to a certain sector. Whereas electricity is something that affects everybody. Everyone’s using it for a lot of different purposes, and so it has a broad impact in the economy. And then the other thing it does is, it’s kind of like a base layer that enables a bunch of complimentary innovation. So again, with the example of electricity, because you have electricity, which is a much more evenly distributed power source than steam, which is about some giant engine at the middle of a factory, you can reorganize the factory floor to take advantage of a distributed power source. You can electrify homes so that people can plug in devices and have appliances. So all those things are kind of downstream of the original innovation. And those things have unlocked huge gains in the structure of jobs that take time to diffuse but are ultimately really transformative. And so the question is, is AI on that scale and how would we know?

Kerr: And coming to this recent period, I’m curious your take on how to interpret the timescales, because you have the famous Bob Solow quote of computers are everywhere except we don’t see them in statistics and so forth, and people talked about electricity just like you’re describing how long it took for us to observe it. With the change that you’re already seeing—and generative AI would’ve come at the very end of that kind of time period that sits in the 2017 to 2024—is it just that it is already signaling something very big that’s going to happen in the future, or is it kind of something that’s more of the digital foundations that we’re allowing for generative AI to start to take root? How have you thought about that? Is this a premonition likely for something that’s going to be much bigger if we forecast out another 10 years?

Deming: So I think there’s a couple of ways to do this. One is, you can ask, “What is the adoption rate of generative AI now, roughly two, two-and-a-half years after the release of ChatGPT, and how does it compare to the speed of adoption of other major technologies?” And so, in a paper that I’ve written with Alex Bick and Adam Blandin, we did this nationally representative survey of generative AI usage, and we constructed the survey to be nationally representative and to mimic the current population survey [CPS], which is a very widely used labor market survey, because the CPS in the past has asked questions about the adoption of personal computers and the internet. And so we kind of put all those things on the same scale and tried to ask the question, “Is the penetration rate of AI at this point in time, relative to its mass market adoption, is it on the same scale as these two technologies?” And the answer is, basically, “Yes.” Right now, about one in four workers say they’ve used generative AI at least once in the last week at work. The penetration rate of personal computers was something like 20 percent at this same stage, and so AI is about 25 percent, and usage outside of work is much higher for generative AI than it was for personal computers and for the internet. It doesn’t mean that it’s as big of a deal on some grander scale, because again, why is the penetration rate of AI so high? Well, it’s pretty cheap or free if you don’t have a subscription, and it’s easy to use because everyone has a computer and everyone has the internet. And so it’s directly building on these two technologies, so it’s maybe not a fair comparison. But the way I think of it is, if it’s already in this many places and in the hands of this many people two or three years out, history tells us it’s not going away, and it’s probably going to become a bigger deal, and we’re probably going to see it be like the way computers are today, which is that they’re everywhere.

Kerr: Yeah, if I recall correctly, too, that study found that younger adoption rate, that young adults had adopted it at a much higher rate than older adults, and so that’s going to be some also signal of what lies ahead.

Deming: Yes, that’s right. And a lot of these things, the way they work is that it’s the younger generation who makes these changes permanent, because we’re raising AI natives now, and older guys like you and me are kind of futzing with it but are not going to fully adapt our routines. But the younger generation is going to really do it. So yeah, that’s what history tells us. We can try to buck the trend, though, Bill.

Kerr: We’ll try our best. A trend that got bucked, when I was reading the paper again with Larry and Chris that you wrote, I was learning that what we call the “polarization” of the job market had come to an end. And so tell us a little bit about that history and then what you’re kind of capturing it for the most recent years.

Deming: So polarization is a trend that’s been happening for a couple of decades now, which you could think of as barbell-shaped employment growth. So if you arrange jobs—from jobs that pay minimum wage or close to it, so low paying jobs, and then jobs that pay, let’s say, a middle -class income but not that much, and then high-paying jobs—so you just kind of arrange all those jobs from left to right and then you plot the growth of employment on those, what you see is growth at the bottom—so low-paying jobs, usually service-sector jobs—growing fast; high paying jobs—like managers, doctors, lawyers—growing fast; and then middle-class jobs—like unionized manufacturing or relatively high-skilled clerical work—those jobs shrinking. And a bunch of studies have found that was happening starting in the 1980s, not only in the U.S. but in many advanced economies. So this is kind of a global phenomenon. And so what we show in this paper is that polarization is over, in the sense that the U.S. economy is no longer polarizing. It’s actually more like an upward ramp. So the low-paying, mostly service sector jobs, are shrinking. The middle-paying jobs are also shrinking, but not by as much. And the high-paid jobs are growing very rapidly. So it’s more like upgrading than it is polarization. And that’s post-pandemic.

Kerr: And, David, to continue on that, it could matter a lot for someone, especially at the lower end of the labor market scale, whether that was happening within person or whether that was happening across cohorts over time. So do we have a sense of that? Is it that more people are somehow being pushed out of the labor market at the lower end, or that we’re actually being able to accomplish things like reskilling that’s helping people move to better opportunities?

Deming: There’s all kinds of these things happening for individuals. But broadly, I think what you’re seeing is that in the ’90s, especially, we had a kind of relative oversupply of—I don’t want to say, low-skilled sounds pejorative—people who don’t have any specialized training for any particular job seeking employment. And so there was a lot of slack in the labor market in the 1990s and 2000s. And so if you posted a job for retail sales clerk at some store in a mall, it was easy to find somebody. And then what happened was—some of these are demographic changes, some of them are the fact that the college-going rate in the U.S. has actually gone up a lot in the last 15 years. So you’ve got more people who are capable of filling more specialized jobs, and then a relative scarcity of low-skilled labor or uneducated labor, which is bidding up those wages. So, actually, what you’re seeing is huge wage gains since Covid at the bottom. So huge wage gains for people who don’t have college degrees for people in often manual labor service-sector jobs. And, so when employers say, “Oh, well, I can’t pay somebody $7.50 cents an hour to work at the retail store in the mall. Now I have to pay them $15 an hour,” you hire fewer of them.

Kerr: Because they’re more expensive, right.

Deming: And so on that, what you see is just fewer of those jobs, because employers are creating fewer of them because they have to pay so much.

Kerr: It’s a fascinating take. Usually, when people have talked about the decline of the middle-skills kind of workforce, it’s often been said, “Well, computers came in, and they could do all the back-office stuff. We got more efficient.” But you’re highlighting those—both the technology, but also that demographic story of a group—that there’s a lot of potential labor supply, and that has now been exhausted.

Deming: Yeah, and I think we just had a couple of decades of relatively plentiful cheap labor, and that has come to an end, and you see that pretty clearly in the data.

Kerr: So you think about artificial intelligence as often having a competitive edge and this must-have skill. Do you have any thoughts about practical ways it’s going to be happening in the labor market over the years ahead, white-collar professionals, and are their jobs going to get displaced by this? Any early clues as to way that’s going to take shape?

Deming: I, like you, probably, have heard a lot of anecdotes about a white-collar recession, big firms not hiring, especially not hiring entry-level grads, difficulty of college grads, MBA grads finding jobs. I haven’t seen a lot of hard evidence on this in the data, but the data sometimes lag reality. So it’ll be very interesting to see if that happens, because it’s something that’s actually consistent in some ways with what we think AI can do, which is a lot of, again, entry-level white-collar knowledge work. And in a lot of the studies of generative AI where, let’s say, you bring in a bunch of people and you give a random half of them access to ChatGPT, it does tend to help the lowest performers more than the experts. So if you’re using it to write ad copy or customer service, it’s really going to help people who are non-native speakers, people who don’t have a lot of context, people who are not experts, let’s say, write software code in a different language than what they know or something like that. But if you already know a lot of things, it’s not helping you as much. And so people tell that as a possibly hopeful story about inequality. But what it suggests is that maybe you don’t still need the experts, you don’t need the entry-level workers as much. And so that’s kind of consistent with a macro story of a white-collar recession in some of these jobs. So conceptually, the pieces fit together. But as a cautious, “This time’s probably not different” economist, I’m waiting to see some hard data from a survey like the CPS before I’m ready to declare a trend like that.

Kerr: Yeah, yeah. It’s fascinating, because in part, you hear a lot of this happening at the hiring margin, people see that they’re not getting a job that they wanted to, and they’re kind of recognizing it. But I’m just fascinated, because also to go back to your existing study, which talked about how younger workers are the ones more likely to use the generative AI and to use the very technologically forward products. And so the story has both them missing out on opportunities, but also being the ones using the jobs or using it at work in a more coherent manner.

Deming: Yeah. And so it’ll be interesting to see how this shakes out. But it’s very difficult to go from anecdote to the data to understand the impact because there’s all these second order implications of it. For example, this thing of like, “Oh, it helps people who are lower on the skill spectrum.” That sounds like a good story, except that if employers internalize that, then it changes who they hire. And so there’s all these second- and third-order knock-on effects that are going to take a little while to play out.

Kerr: So we talked about at the beginning how this report has been very widely read and talked about. What’s been the discussion like? And has there been any particular reactions or threads that have emerged that were surprising to you, something you hadn’t anticipated?

Deming: Well, I got to say the analysis we did at the end was really surprising to me. I had not expected to find… In some sense, when we obviously wrote this, we thought we were writing one paper, and then we ended up having to tack on a second paper at the end, which was things are changing a bit faster. What I will say is that sometimes I feel like, as labor economists, we’re tracking the recent past, and sometimes you write a paper just documenting a trend just as it’s reversing. We wrote this. One of the findings in the paper is that we’ve seen a huge run-up in the last decade, basically from 2012 onward, in hiring in STEM fields, science, technology, and in management. So the exact number is that the employment share for STEM jobs went from 6.5 percent In 2012 to about 10 percent in 2022, 2023. So that doesn’t sound like a lot, three-and-a-half percentage points, but relative to the base rate, it’s an increase of about 50 percent. So it’s a huge increase in STEM hiring. And that’s driven, I suspect, by big -tech companies hiring both software engineers, software-type jobs, but also what we see in the data is they’re business-type jobs that are data intensive. So huge investments in data centers, huge investments in AI-related technologies, hiring technical talent. Some of it coming from overseas, but a lot of it hiring native-born graduates. Computer science is now by far the most popular major in the U.S., so a huge run-up in CS talent. Yet, when you read the news, what you see is the opposite, which is hiring for software developers has cratered in the last couple of years. And so I just wonder whether that’s actually things have changed or whether it’s just people overshot the mark, and now they’re scaling back, but actually the trend is going to continue once it adjusts. So I sometimes feel like we’re always chasing the story a half-step too late because of lags in data availability and things like that.

Kerr: Yeah. Well, one place that I think you have been on the story for a long time—but let me bring it into context with the STEM increase, which I believe is happening and probably will continue—is the work on social skills. And you have some very famous work around the rise of social skills in the workplace and their need. Do you anticipate that bundle going forward, that the optimal bundle as being one that has the digital fluency that comes with kind of a STEM background and the social skills being married together?

Deming: Yes, I do. And I think that the social skills are going to become even more important. And I think part of it is related to the ability to work in teams, which is what my recent research that you so kindly described was about. But there’s another aspect to it, which is the ability to form relationships, build trust, and make meaning with others, I think, is going to become more important. And I think that’ll be especially true if we make more rapid and more serious advances in AI. So let’s just grant the premise that OpenAI and companies like it achieve their goal of creating artificial general intelligence. Intelligence isn’t that well defined, but let’s just say when we created a technology that can do most workplace tasks more intelligently than everybody. So if you grant that premise, it doesn’t necessarily follow that it’s going to replace all of our jobs, because not all jobs require intelligence as their main input. Intelligence helps in almost every job, but there are a lot of jobs where it’s not even the primary thing that you want, and a lot of those jobs are jobs where forming relationships really matter. So if you think about coaching, mentoring, guiding people, which is a surprisingly high share of jobs. I gave this example, I gave this talk at Junior Family Weekend at Harvard when families were asking questions about are robots going to take all of our jobs, and I just kind of came up with this off-the-cuff example that I, on the side, coached my daughter’s soccer team, and I don’t know anything about soccer. They joke with me, they call me Ted Lasso. I don’t know nothing. Clearly, you could find someone to be a more intelligent soccer coach. Maybe I could find an AI coach that optimizes substitution patterns and figures out strategy and formations. But nobody wants that, nobody’s asking for that, because they’re not asking me to be a coach because I’m great, I’m the most intelligent coach. It’s because they want me to form a relationship with their child, be able to be a trusted resource, have some accountability for things that happen. And those are all, I would say, just intrinsically human things. You actually want the person because of what it means to have a person there. So I foresee an economy where building trust, establishing a reputation, we will have a luxury services economy, where you could get the generic AI commodity, but you’re willing to pay extra for the person that’s more scarce. That’s not going to happen tomorrow, but I see that as the trend, and that’s a world where social skills are the most important thing.

Kerr: At the recent American Economic Association meetings, I attended one of the panels that had a bunch of leading figures in AI kind of contemplating what the future could look like. And one of the panelists gave a comment, and it certainly gave me pause, that for his children, they’re going to grow up in a world where, basically, the machine is always going to be smarter than them, at least across the broad range of human activities, whether it’s ChatGPT or whatever, AGI future exists. The second part of it was he definitely thought they should at least finish high school. He wasn’t very clear as to whether further training into college was going to be very useful or important. Do you think that college’s value is going to go up or go down or just stay the same in the future? Are we going to educate as many college graduates going forward as we have been ramping up over the last few years?

Deming: I think it’ll go up. I do. For a couple of reasons. So if you look at the share of jobs that “require” a degree, it’s pretty small, but it helps in an awful lot of jobs. And I’m willing to bet, even if you have a plumber, you hire a plumber, I bet the plumbers with bachelor’s degrees are better, and they’re certainly going to be better at things like customer service, understanding and anticipating people’s needs, communicating with a fellow bachelor’s degree holder, all those things are still valuable. And so I think we’re going to want that. What you’ll see is an upgrading of people’s expectations for the human work that we’re doing, and I think education will probably have to change to accommodate that. So I think you would expect to see, as you’ve already seen, less focus on memorization and rote comprehension and even synthesis of information, and more focus on the higher order things, things like working together with people with different perspectives, talking across difference, things that are more valuable when you have an abundant intelligence available to you. But again, we want more education to be better at these jobs, even if it’s not strictly required. And education, like healthcare, is one of those things that when we get richer and more prosperous, we want more of it. So we’ll live longer because of medical advances in AI, we’ll have more time to educate ourselves, we’ll want to be healthier. And so I would go long on education and healthcare in terms of their importance in the economy over time for these reasons.

Kerr: Yeah. I guess on maybe a final just broad question on public policy. Given what we have observed with your data, first few years, it is still, as you correctly highlight, it’s hard to predict what it’s going to be like in 10 years. Is there anything that you would be next thinking about on a public policy lens, relative to occupational turnover and jobs and the like?

Deming: Yeah. So I would say that I’m skeptical of any attempts to regulate the form of the technology right now, not because I think it should never be regulated, because I think it’s just far too soon. I think if you just even look at things like the California regulation about AI based on compute size, and then DeepSeek comes out. There’s too many unintended consequences to try to directly regulate the technology. Having said that, I think it’s important to try to foresee some of the issues that the technology will create. I don’t think it’s going to take all of our jobs, so I don’t think we need UBI [universal basic income], but it may lead to, let’s say, some disruption of white-collar office work, and so we may want to think about policies that help people relocate or retrain for other types of jobs, maybe for blue-collar jobs. So I think those kinds of things, which are downstream of the technology, we should absolutely try to foresee in terms of policy. So I think that’s important to do. I would draw a distinction between that and directly regulating the technology, itself. Maybe when the technology’s more mature, it might make more sense to put some guardrails around it, but I think it’s too soon right now.

Kerr: All right, David, final question is, you’re very active in the research front. What’s next around this nexus of generative AI, the labor market? Where do you want to go next?

Deming: Yeah, so a couple of things. One is on the kind of macro side, I’m very interested in this idea of a white-collar recession. I’m looking at could I possibly write a paper investigating that, trying to do something more systematic with it? So that’s kind of on the macro side. Don’t know if it’s going to happen, but I’m alerted to the possibility. And then I’ve been doing a lot of work on soft skills, as you mentioned earlier, and I’m really interested in the skill of working with AI, in particular AI as a decision-making aid, in a world that’s uncertain and you have to make decisions about different paths to take and what are the conditions under which AI will hurt or help, and who are the people who are good at working with it. So I’m doing a lot of work on AI-assisted decision making that I’m excited about.

Kerr: Well, I’m sure both of those, when they come out, could get a lot of attention as the work we’ve talked about today. David, thanks so much for joining us again.

Deming: Oh, it was a real pleasure. Thanks so much for having me, Bill.

Kerr:We hope you enjoy the Managing the Future of Work podcast. If you haven’t already, please subscribe and rate the show wherever you get your podcasts. You can find out more about the Managing the Future of Work Project at our website hbs.edu/managingthefutureofwork. While you’re there, sign up for our newsletter.