Mike Spence: What a Nobel Laureate Thinks AI Does to Education

Ben: Signaling is a hugely important part of every labor market interaction, but it wasn't well understood until my guest today introduced the idea 50 years ago. Now the value of signals is changing rapidly, and he's thinking carefully about what comes next. Mike Spence is a Nobel laureate in economics and former dean at both Harvard and Stanford. His work on information asymmetry and growth is foundational to how economists think about markets and development. Here's our talk.

Ben: Welcome to The Economics of Work. I'm honored to welcome Mike Spence, Nobel laureate and former dean at both Harvard and Stanford, and just a really clear thinker and writer on so many important topics. Mike, welcome to the show.

Mike: Well, it's wonderful to be with you. Thank you.

Ben: Thank you. I want to start with labor market signaling, which is where you first really made your mark, and what you won the Nobel Prize for 25 years ago. So happy 25th anniversary.

Mike: Yeah.

00:53 Is AI eroding labor market signals?

Ben: So in a world with open access to AI, where anyone can generate really high-quality, tailored content, are the labor market signals that have previously been delivered, like job applications and cover letters, eroding? And if so, where does that take us?

Mike: There are various dimensions. This is a really interesting question. Let me give the short answer: yes, they are eroding. We have these powerful machines that can imitate anything you want them to imitate, so whatever subtle signals were delivered in CVs and other things people submitted are going to be muted, meaning they're going to lose their informational content as the market naturally evolves. There was a kind of flamboyant article on this general subject. You know more about this than I do, but the basic idea is that there's a flood of CVs coming in, most of them generated by AI, and on the employer side there's a flood of software designed to weed out all that stuff. So you have a CV screening war where nothing of a traditional kind emerges anyway. But markets generally evolve, and there will be a process of looking for signals that don't lose their informational content over time. What exactly those will be is a different question. But I would not want to be on the side of a debate that said this isn't going to have any impact on the informational structure of the labor market.

02:52 Can AI matching reduce information asymmetry?

Ben: I could see two contradictory paths. One possibility is that we get completely signal-jammed, people default to other forms of signaling, like networks, and we just have less information in this market. Another possibility is that you have these AI-generated applications and AI-generated screeners, but they're really good. They can capture a lot of the dimensionality of someone's character and the nature of a job. Let's say the AIs get so good that they can peer into your soul and capture the very essence of who you are as a person, and they understand the essence of a company's job requirements the same way. Through this AI matching, we get better matches that capture more dimensionality than a traditional cover letter ever could. So could this reduce information asymmetry in a way? Is that plausible?

Mike: Well, it's definitely plausible. If you step back for a second, the internet itself, before we had any sense of the breakthroughs coming in AI, dramatically changed the informational structure of markets, because it took a whole lot of information that was either private or costly to get and made it so cheap that you could get it. There are lots of simple examples. When you bought a car before the internet came along, there was basically nothing you could do other than listen to the salesman, used or new, and try to guess if the guy was being honest with you or not. Now, whether it's cars or medical treatment or almost anything, there's a massive amount of information. You still have to be careful about which information you use and which you don't. So even before you get to AI, you've got a tremendous change in what information is available and in its implications for market behavior. And we've accepted all that. By and large, I think it's been beneficial.

With AI, I think you do have powerful tools that, used properly, could solve matching problems that are very difficult to solve otherwise. Al Roth has thought deeply about these things-

Ben: Former guest on the pod.

Mike: He has this wonderful new book. There's a lot going on in addition to solving the matching problem, having to do with incentives and even norms and values. There are lots of cases where people really don't want the market to function, because there's something wrong with that idea. But on the technology itself, I think you have a reasonable basis for thinking that its ability to keep track of variables at a level human beings are incapable of might very well produce matching and related processes that are much, much more precise and efficient.

The only qualification is that every time somebody looks at these things, there's the technical side, what can you do, but then there's the incentive side. Milton Friedman and a whole lot of other people kept reminding us: it's fine to have a policy that sounds like it works, but if you haven't paid attention to the incentives of the players in the game, you could get blindsided by outcomes you didn't expect.

Let me give you an example. It's not the labor market or a generalized matching problem, but AI is being used pretty effectively, on an expanding basis, in supply chains. One of the reasons is that supply chains are incredibly complicated, and the AIs have a way of figuring out how they're configured and how to reconfigure them when a change comes. At least among the people I've talked to, there's a growing body of evidence that this is really quite effective, meaning there's a big difference in both the speed and the effectiveness of the response. And the AIs become more valuable the more volatile the environment is, right? Because the more volatile the environment, the faster you have to make changes, and the more of them, relative to a very stable environment in which you slowly figure out the answer without the benefit of the AI. So on multiple dimensions, my response to your speculation in that direction, which is probably more than speculation, is definitely positive.

Ben: I really like this framing of AI as a continuation of the internet, where we entered this information age with tons of information. I don't know if this is an overstep, but we can think of AI as an information synthesizer. Standing on one foot, that's what it's about: synthesizing information, which lets us analyze really complex supply chain information that used to just sit in some database, super hard to analyze. So it makes sense that we're finally living up to the potential of the information age. I also wonder-

Mike: You can think of an early version of AI, depending on how you count, as the evolution of the search engine, right? People tend to separate those things, but there's quite a lot of artificial intelligence built into search engines these days. They didn't call it that when they first started out on this journey. That was a major, major step in taking this mass of undigested information that suddenly got digitized, where you had no idea what to do with it, and making it accessible. So there's an element of continuity in this. Not to understate the power and differentiation of these modern AI models, but I do think there's an element of that in the digital age.

09:58 Do we still need costly signals like education?

Ben: Interesting. So let's say we'll be in a world with more information, and better understood information. Do we still need signals that are differentially costly? That doesn't really help with signals that have differential costs, like education, where, as you pointed out 50 years ago or so, the value of education is that it's less costly for someone who is high ability, high potential, and more costly for someone who would have a harder time. So where does that leave education? Is there still a role for these signals that are costly by design?

10:47 Signaling vs. human capital: the Chicago debate

Mike: Well, we had a temporary disruption early on, when the people in Chicago thought that signaling was an attack on human capital theory. I had to go to Chicago, and we had a seminar. It set the record straight. I said, "Well, this isn't in competition with human capital." I don't have any doubt that in educational institutions a fair amount of human capital is added, and a signaling effect-

Ben: Just out of curiosity on that point, who took it most personally? Was it Becker and Murphy, or others?

Mike: No, it was just a misunderstanding. I know where it came from, right? Tom Schelling told me, "Well, if you want people to pay attention to this, then write down the simplest model you can where there's a signaling effect." And by simplest, he meant that if you take one additional feature of structure away, the signaling effect goes away. It just won't be there. So I did. It was natural to get rid of human capital altogether. Actually, that's not the first thing I did. I had a model with human capital, in continuous variables, with differential equations and so on, and he said, "They're not going to get that." So the first signaling model anybody saw had two groups and one signal, with no human capital and a whole bunch of other stuff stripped away. So they thought, oh gosh, there's this young guy proposing an alternative theory of where value is either created or redistributed in higher education. That was the origin of it. When I pointed out that it was very easy to write down a model in which you had human capital and signaling at the same time, subject to empirical verification, of course, that went away.

Ben: I get that they can entirely coexist, of course. But given that there's a signaling component that's presumably independent from education, it would stand to reason that education is, in a way, oversupplied relative to what's socially optimal, right? That sounds fair to say?

Mike: Well, relative to what would happen in a world of perfect information, that's correct.

Ben: So if we're tending toward a world of more perfect information, would the gross value of education decline on the margin because the signaling value declines? Or will universities and other education providers need to do more education relative to credentialing? How do you see this interacting with what educational institutions do, and the role they have?

Mike: I don't think this is likely, at least in the short to medium term, to produce a dramatic change. Will the graduates of Caltech suddenly look like the graduates of a community college? Probably not. So I think there's something real going on under this. Maybe it occurs at the admissions office, as opposed to in the human capital that gets built in the two or four years you spend in the institution. But when it really matters, there is going to be a need to find a way to get valuable information, in the case of the job market, about which people look like they're high potential. And people will find a way to do it.

15:00 Testing students when AI can write the exam

Mike: In education, and this is slightly aside, there's a lot of hand-wringing, as you know, about how we're going to test people. And I thought to myself, okay, so you can't have a take-home exam anymore, because somebody else, called an AI, is going to write it. But this isn't an impossible problem to solve. It's a little bit labor intensive, but just sit in a room with somebody, with no technology at all, and have a conversation about the subject. Maybe some colleagues of mine don't want to hear that. But to some extent, I think the answer is that we adapt and find ways to learn what we really need to know. That example is outside the labor market, but it's maybe related in terms of process. I don't think we're dead in the water when it comes to assessing whether people have really advanced cognitively in various fields.

Ben: Yeah, I think that's right. I do that in my course. Up until a couple of years ago I was doing take-home exams, and now I do blue books in the classroom, and it's fine. It's not a big deal.

16:29 Grade inflation and elite college anxiety

Ben: So there's an easy workaround in that case. But that brings up another issue I've been really curious about, which isn't related to AI but seems to be happening at the same time: there's a big trend in grade inflation, which also seems to be signal-eroding. People have been talking about grade inflation for a long time, probably also when you were a dean. Why is there meaningful grade inflation now, and what does it do to labor market signals?

Mike: That's a hard question to answer, at least for me, in the sense of having a comprehensive understanding of it. Whenever I confront something like that, I ask myself: for whatever reasons, and they can be pretty complex, have the incentives changed? Is it harder, if you're a faculty member now, to stick to a predetermined grade distribution than it was, say, 20 years ago, when yes, we had grade inflation, but it wasn't as extreme as the current version? And if you think that might be part of the explanation, then what's going on?

This is the point where it gets a little complicated. In American society in particular, there seems to be this feeling that if you don't graduate from an elite institution, one whose name you can write down on a list of 20 or so, then you're out of the game. When I look at that, I think it's crazy. One of the hallmarks of American higher education is how many very, very good institutions there are, how many options, and how much product differentiation. You have teaching colleges that are absolutely high ranked, and people leading major organizations who graduated from them. You just don't need an Ivy League degree or a Stanford degree.

So the question is, where does that come from? I don't pretend to have an answer, but it seems to me that we have somehow generated, through a combination of beliefs and social structures, a set of incentives that are misguided, as I would put it.

Ben: Yeah.

Mike: The extreme form of this is that we're starting to prepare our kids for college before they're out of elementary school. Maybe I jump to it too quickly, but I get off the boat at that point. That's a terrible idea, right? Kids should be kids and follow their passions, rather than getting prepared for something. There's a deeper conversation around this, which is: how do people thrive? Do they thrive by wanting to be something 20 years from now? Or do they thrive by being what many scientists are, which is people who are fascinated by stuff and get up in the morning because they want to get on with figuring it out? But why this is happening now, I really don't know. Maybe it's a set of values that equates, or partially equates, accomplishment with wealth. I honestly don't know.

Ben: It's very tricky. I struggle with this too. In my own classes, I do feel the pressure for grade inflation. It's a real pressure. I can't speak to what it would have felt like 20 years ago, but I've been teaching in universities for 15 years or so, and I feel like there's a little more pressure now. It's hard for me to pin down where it's coming from.

Mike: Maybe it is just a vibe shift. If you've got a B or a B-minus in a bunch of courses, is there a feeling among your students that they've just eliminated a whole bunch of options?

Ben: Yeah, I think so. There are also formal constraints. Students say, "If I get below a B, I lose my scholarship," or whatever. So there are some institutionalized incentives, which are probably distortionary, but I feel like it goes deeper than that.

21:55 Would a national exam be better?

Ben: But let me ask you this. Let's say we forget about distributions within universities and give up on that whole concept. What if we had a system like AP exams, with a national exam for every course or course type and a fixed distribution? You'd expect people who were selected into elite universities to get better grades on average than people at less selective ones, but it would all be one global distribution. Putting aside who's incentivized to support that program, let's say you're the emperor of the world and can just pass it as a policy. Would that be a better world, in your opinion?

Mike: It might be. One wants to be careful. Would it supplant the existing systems? I don't think so, because the underpinnings of a lot of those systems are: which network are you in? And people know, correctly, that that matters, right? Not that it's immutably determined by what college you went to, but those things are really powerful and important. So I don't think a national exam would supplant that, but it might create opportunities for a subset of students that are harder to access otherwise.

In China, for example, they have a national exam for entry into the universities. It has a long history; they used to have exams for entry into the civil service. And I think there's a legitimate element of meritocracy in that. Now, it may produce kids whose parents have them studying 16 hours a day to make sure they come out okay in that system, so it won't necessarily produce totally benign results. But with a national exam, an outstanding kid who, for whatever path-dependent reasons, didn't end up in a traditional high-prestige institution could show up in a way that would be harder to detect otherwise.

Ben: That sounds interesting to me, though I would think it would eat into the value that elite institutions provide. Of course, they have the benefit of the best educators and an elite network, which have real, causal, additive value. But here's one thing I like about it. One critique of Harvard and Stanford and the best schools is that they don't educate that many people. They're selective by design, and that selectivity is part of the signaling value. But if that goes away a little bit, and people become a bit more indifferent between going to an elite school and an average school, then I would think the universities that are best at education would do more education. That seems really positive.

Mike: I agree with that. And to support your point, based on the data I've seen over the 20-plus years since I stopped being an academic administrator, the selectivity numbers are in the ridiculous category. These schools used to be selective, but not at this level. And that has produced more anxiety, I think. There used to be a 5% chance of getting into one of these places, and now it's one and a half percent or something like that.

Ben: Yeah.

Mike: On the other hand, unless these schools are so well endowed that they can basically absorb the high and rising costs, eventually people will start to say, "It would have been nice, but we can't afford it, so we're going somewhere else." And maybe that's okay. Then eventually the employing side will say, "There are a lot of pretty talented people who aren't going to those institutions anymore. Why don't we go take a look at them?" So it's hard to know how this will come out, but there are adjustment mechanisms. I do think the cost of higher education is another dimension in which serious work needs to be done. Maybe these new tools will create either competitors or opportunities to reduce costs without diminishing the effectiveness of the activity.

But on networks: if you live and work in Silicon Valley and somebody says there's a perfectly good alternative in Ohio, nobody's going to listen, because those networks are real. My friend Jim Gibbons, who used to be dean of the engineering school at Stanford, said it's a body contact sport.

28:05 Learning outside the classroom

Ben: Yeah, that tracks. In some ways, I feel like that exists even apart from the education. One topic we discussed with Rachel Lipson in a previous episode, when she was running the Harvard Project on Workforce: she does a lot of work on community colleges, and community colleges really don't have that type of network. Maybe they provide a great education, but they're commuter schools; they don't really have that. We contrasted four-year colleges with campuses to summer camp, in a way, where people hang out with each other and you learn a lot just from socializing and having random conversations. A lot of education probably happens outside the classroom, which seems quite valuable in its own right.

Mike: I completely agree with that. The surveys of the graduates of these places confirm it. If you take Stanford Business School and ask people what the most valuable part of their education was, the answer is: my classmates. It's that blunt. It's probably an overstatement, because you need the context; they're not just milling around. But I think that's a very important point.

I have a more radical view, which is that learning doesn't even occur in the classroom. When I started teaching, I was dreadful, because I thought I had to cover a whole lot of material and that I was transmitting information or knowledge. Then it finally occurred to me that maybe that's not what was going on. I was supposed to be getting students interested in things, and their learning process was occurring, at least in part, outside the classroom, as they processed the information. I remember taking Kenneth Arrow's general equilibrium course at Harvard as a PhD student. There's no chance of absorbing general equilibrium theory in its entirety in a classroom.

Ben: Yeah.

Mike: The way we learned it was that everybody went home and tried to write the models down on their own until we could figure out how to do it. That's why the interactive part of teaching, which has to do with motivation and curiosity, is not going to be written out of the script. It'll be a complement to these powerful tools people use. In the context of education, I say AI is the most powerful digital assistant slash tutor you're ever going to have, because it's read everything, in every language, in every field, and at some level understands it. You're not going to find a human being who can do that. So use it wisely and well. Don't get me started on this, but when I do this in a classroom, I say: look, you're going to use this, and it's really useful. But when you use it, disclose it, and never use it as a substitute for yourself, ever. That's principle number one. Because if you do that, you're abandoning your stated goal for being here, which is your own personal cognitive and other development.

31:43 Will AI make us mentally unfit?

Ben: That reminds me of a framing I sometimes use, comparing it to exercise. Let's take the premise that we now live in a world where we don't really need to lift heavy things or move our bodies very much. We used to, but we don't anymore, so there's no requirement to be in physical shape to succeed economically. As a result, or at least coinciding with it, a lot of people have become very unhealthy. There are tons of people in the world who are far too unhealthy. But then there's some class of people who exercise, and I think that is a signal. I'm not even sure what it's signaling, but it's some sort of fitness for the world: that you have the wherewithal and the discipline to exercise and take care of yourself, even though you don't need to economically.

I wonder if there's a parallel with intelligence, or mental fitness. In a world where everyone has a super genius on their shoulder, in this world of AI companions, maybe intelligence in its own right will be less important, the way physical prowess is less important economically. And maybe some people will fall off the wagon and become very stunted mentally, and we'll need a new kind of signal, some exercise equivalent, to show that we're mentally fit. What do you think of this model of a future where we don't need intelligence the way we used to? First of all, do you buy the premise? And second, do we need something like exercise, where we do things the hard way even though we don't need to?

To make it a little clearer: when we exercise, obviously a machine can lift weights, but we don't have a machine lift our weights for us, because we need to do it the hard way to develop. In school, we can have AI do our assignments, but then we're not going to develop mentally. And maybe that's okay to the extent that it's not visible, and we can fake our way through life with this mental crutch. So I want to gauge your thoughts on that mental model: physical versus mental crutches.

35:01 Human-AI collaboration and the Turing trap

Mike: It's a really interesting set of questions. I'm not sure I have a fully developed answer, but let me start here. There's a great deal of talk about the automation of this and that. Erik Brynjolfsson talked about the Turing trap, and now people are terrified of AI agents. I'll get to a direct answer to your question, but my view is that for the foreseeable future, the right model is mostly not automation. It's machine-human collaboration.

Let me give examples. I have a friend who's an expert on financial infrastructure. He helps countries that don't have any financial infrastructure figure out what they need to do. He uses very powerful agents to help him do that work now, and it saves a massive amount of time. But there are probably six agents producing bits and pieces of the thing, so the integration is still important. Maybe an agent will be able to integrate it, but it's still a proposal. I can put it several ways. One, they still make mistakes. Two, there's only one person who can fix it, and that's the human who understands it and looks at it. Three, it's more like a first draft, right? Anybody who thinks it's the final draft and hands it in is nuts. It's just a mistake.

Eventually you may get to the point where it hasn't made a mistake for the last two years. So I don't dismiss the notion that we could lose some mental muscle tone over time because we don't have to do certain things. But then I go back to education, Ben, and I say: do you really want to be unfit, mentally and physically? In the end, this is a question of values more than anything else. It's about what people think is important.

There's a legitimate amount of talk, angst, and analysis around the impact of these things on work and labor markets. But there are hugely important impacts in places like the sciences as well, as you know very well. Scientists view these as incredibly powerful tools that help them get things done that either took a huge amount of time or couldn't be done at all. So there's a legitimate element of excitement in that world. There are people who have stood up and said AI will design the experiments and generate new knowledge with no human intervention. Maybe, but I don't honestly expect to be around to see that happen.

38:29 Is the real risk automation or a productivity surge?

Mike: For the time being, I think the dominant useful model is that, yes, there is some element of automation. But let me put it differently, because you're a labor economist. If I ask myself whether the real problem is automation, or a massive productivity surge that produces a loss of labor demand because the demand side hasn't kept up, I would normally pick the second. I can see the productivity effect coming. And I don't think people care whether they lost their job to some useful or misguided bit of automation, or because the short-to-medium-term general equilibrium outcome was that the productivity surge wasn't matched by a surge in demand.

Everybody knows their jobs are going to change and they'll have to move around; there's a general consensus on that point. But that doesn't answer the question: are we going to have an employment problem? Once we figure this out, I don't think it's full-throated automation we really need to worry about. It's a massive shift in productivity in various areas, not tomorrow, but over a short enough period and coming fast enough. I don't know if that's going to happen, but when I worry about the employment problem, that's the direction my thinking has gone.

Ben: Interesting. I want to drill into that, but first I want to say that I think we can believe in complementarity and avoiding the Turing trap, and also believe in some kind of mental decline. I liked your teaching example: you have an AI that has the intelligence but not the motivation. It can't motivate you or help with the more human, relational things. When our muscles were automated in the first industrial revolution, the first machine age, let's say, then, to quote Erik Brynjolfsson, we relied more on our mental capacity and let our physical capacity go. That's fine, because mental capacity is also really important. Maybe in this next machine age, we'll rely more on our relational capacity, which sounds nice and positive and complementary with AI. But maybe we'll have some mental decline, the way we had physical decline. So we can believe in that without necessarily believing in automation or labor displacement.

Mike: I accept all that completely. If you don't have to dig trenches and you can get in your car to go to work, then, averaging over several industrial revolutions, there's not only a risk but probably a near certainty of some loss of physical fitness. That raises two questions. One is, do you accept that as inevitable, or do you try to limit it in various ways by shifting the incentives? I live in Europe. People drive cars too much there as well, but most European cities are configured in such a way that you don't have to. They're also older. They were built before cars. They're hopeless with cars. My wife and I never use the car in Milan at all-

Ben: Sounds amazing.

Mike: Because it's a pain to get anywhere. You have to take circuitous routes, the one-way streets change direction every three blocks, the names of the streets change. But the public transportation is excellent, and you can ride a bicycle and walk. I don't mean to oversell it, but there's a lot of truth in the notion that design features of the environment we live in have an impact. Maybe they're so historically determined and path dependent that there's nothing we can do about them by the time we figure out what's going on. But I think your general point is right: you want to think carefully about the kinds of atrophy these things can produce. My friend Mohamed El-Erian, with whom I wrote a book with Gordon Brown, kept reminding us that when the calculator came out, everybody said nobody would learn arithmetic anymore. There's probably a little element of truth in that, but most people can still do arithmetic. So I'm not sure how this will come out. Ignoring the atrophy issue would be a mistake for sure, but projections that it'll be massive are not necessarily right.

44:31 Will we need fewer software engineers?

Ben: That tracks. So, circling back to the concern you mentioned about a productivity surge, let me see if I understand. We have this big surge in productivity, but demand, not just for labor but for goods and services, doesn't keep pace. So you'd have to believe the economy at large is inelastic, worse than unit elastic. If that's the case in the short to medium term, we'll just be satiated with the goods and services being produced, and there won't be enough labor demand. Is that the model?

Mike: I want to be careful. It is the model, but if you use the word elasticity a couple more times, you'll lose your audience. I've done this.

Ben: All right, I'll take it back. We'll edit it out.

Mike: I talk to people all over the world, and I say: let's ignore automation for a minute, and suppose we have a big productivity surge in software coding. What do you think will happen? With almost no exceptions, people say, "We'll need fewer software engineers and coders." And I say, "Okay, what did you assume when you said that?" They say, "What do you mean?" I say, "Well, think about it. You probably assumed output was constant, right?" And they say, "Oh, yeah."

Then I get around to elasticities. There are goods we call essential goods that have very low elasticities, at least within reasonable ranges of prices and incomes. If you have a big productivity surge in that area, you're probably going to have reduced employment, among other things. But not everything has that characteristic, and there's a fair amount of data, by sector and type of product, suggesting some elasticities are really quite high. So I don't know how it's going to go over time. But to think about it carefully, you need a few simple economic concepts most people aren't equipped with. That's part of the problem we have with the terror and anxiety associated with automation: they're coming for my job. Or the more sophisticated version: even if it's not full automation, if it's a productivity effect, maybe I'll lose my job anyway. And the answer is, it probably depends on where you are in the economy.

Specifically on software engineering, sure, there's disruption all over the place. People who are writing code and developing software using traditional methods, if they stick with them, are probably going to lose market share, not to mention suffer some fairly severe adjustments in valuations in the capital markets. But as a general matter, when I ask people about software engineers, I say: you may be right. But the simple truth is that our economies are becoming capital intensive, much of that capital is digital in one form or another, and a whole lot of it is software. If we're building economies on software, even with AI agents helping us do it better and faster, are you absolutely sure we're going to need fewer people? At that point, I think people realize there are a lot of variables that are very hard to know in advance. A confident prediction on the draconian negative side seems to me not necessarily justified.

Ben: I tend to agree. I heard an example I thought was so funny. The idea is that people think about the economy as if everything's toilet paper. There's a limited amount of toilet paper people need, and short of changing people's diets, you're not going to quadruple the amount of toilet paper that needs to exist in the world. So there really is limited appetite there. But other things just don't have that characteristic.

Mike: I'm going to add toilet paper to my low-elasticity-of-demand catalog. Thank you.

Ben: I thought it was such a funny example. You could give everyone Mexican food, I suppose. But that's one model.

49:54 Keynes, emerging markets, and AI optimism

Ben: This seems to be an especially potent argument when you think about the developing world, and I know you've done a lot of work on emerging markets. If I think back to Keynes's 1930 essay, "Economic Possibilities for Our Grandchildren," which predicted a productivity surge that would leave people needing to work 15 hours a week, I think it has a couple of fatal flaws. One is that it assumes we'll run out of things to do. But it also doesn't account for the fact that there's a whole developing world out there that's nowhere close to satiating its needs.

Mike: Correct.

Ben: What do you think the vibe is in emerging markets? How are they thinking about the threat of AI disruption?

Mike: They may think of it as disruption, but they think it's mostly positive. A couple of observations. You're right, I do spend a fair amount of time out there. There's one misconception we have to get rid of right away, which I was just discussing with a colleague at Stanford. There's still a view out there, not held by everybody, that if you're not in the frontier development of these models in some form, whether you're a country or whatever, you're out of the game. So my first move, almost always, for the subset of people who think that, and it may be a minority, is to get rid of that idea. There are two parts to the revolution. Yes, China and the United States are the epicenter of most of the activity at the frontier level right now. That may not be a permanent condition. But picking these tools up and using them in ways that are not only profitable but advance social, economic, and other goals, and I know we could misuse them too, is a huge opportunity set. With both open and closed models and a fair amount of competition, you have to think about that as an opportunity.

The second observation is about open source in China, even though it may have some problems in terms of security. On the misuse side, if they produced a Mythos and made it generally available, and it could dismantle the vulnerabilities in most operating systems, that would probably be problematic. I don't think they'll do it. They have enough regulatory control to avoid doing something like that.

Two years ago, the Human Development Report was on AI and its potential, and there are surveys of attitudes toward AI in the emerging economies and the developed economies. For the most part, they're uniformly and very significantly more positive in the emerging economies. I think that's because they see the potential for not only job creation but for solving problems that are traditional but important on the development path as pretty significant. Yes, there's an ongoing discussion that Dani Rodrik, others, and I have participated in: is there an alternative model as powerful as the industrialization model that powered many of the Asian economies? The honest answer is maybe, but we don't know what it is, at least at this point. So it's not all positive. But the other important message delivered around the emerging economy world is that you don't want to view this as a silver bullet, but you don't want to miss out either. You need a certain kind of infrastructure to make sure you're in the game, and then you look for opportunities and let the young people, who tend to see them, go after them.

A simple model, and it's not a complete model, Ben, is: make sure you've got the infrastructure you need. That means mobile internet, reasonably reliable connectivity, fast enough that it isn't useless, priced so that it's not restricted to 15% of the population, and so on. Data centers, and in some of the poorer countries, electricity, to be honest with you. That's all crucial. But if you put all that in place and have a reasonable amount of intellectual and entrepreneurial back-and-forth, the evidence suggests the young people will pick it up and start doing really interesting things. You need the financing and all that. But there's a global explosion of entrepreneurial activity around tech, and I don't see any reason why AI won't be part of that, now and increasingly in the future. So there's generally less anxiety and fear about this in the emerging economies, and more of a sense of potential, even though we might not know exactly what that potential is.

56:20 AI and inclusive growth: the dermatologist example

Ben: It sounds very exciting.

Mike: Think about it. In my MBA class at Bocconi, which is about emerging economy growth, growth in general, and what these technologies do, one of the reasons students are so interested is that you can produce inclusive growth patterns using technology that are impossible to produce otherwise. We know this from fintech: digitally enabled financial services reach very low-income or remote populations, who get left out because serving them through traditional channels isn't profitable. But it goes beyond that. You can see it in health and education. People may not have a precise sense of it, but they understand at some level that there's a very powerful inclusive dimension to this, one that may or may not produce massive surges in growth but will produce massive surges in accessibility.

Two students at Stanford, and this is pretty general now, worked with dermatologists to figure out whether you could detect skin cancer from a well-curated, hard-to-assemble set of 120,000 images. It worked pretty well, and I'm sure it's more sophisticated now. But it didn't beat the dermatologists, to go back to Erik Brynjolfsson and the human benchmark. So I say to the class, "What do you think of all this? We haven't beaten the dermatologists, even on detection, let alone pathologies and treatment." And without exception, even the students from emerging economies, and a lot of them are, say, "I'm going to the dermatologist." I say, "That sounds right. I'd probably do the same thing. Anything else?" And they say, "No, that's it. Nice try. It's not at or beyond the human benchmark." And I say, "Okay, but 85% of the world's population doesn't live anywhere near a dermatologist." And they say, "Oh."

Then you start to ask: isn't an imperfect substitute for a dermatologist, as an element of preventive primary care, kind of interesting in parts of the world we don't live in? And they say yes. So they get it, right? There's a bunch of things you could do that give a much expanded array of opportunities and services to people in countries that have a fairly long way to go before they reach developed status, even if they're succeeding, and many of them are. Those things are going to be pretty valuable. And that's a set of opportunities, not only social but entrepreneurial.

59:17 The diffusion challenge

Ben: That's a very resonant example. One theme in a lot of your work, and let me know if I'm reading into this incorrectly, is that there are multiple equilibria and multiple paths, and we have choices to make that can set us on one or the other. So it's interesting hearing you frame this as a path that could be pursued, not an inevitability. What are the choices that have to be made by policymakers, administrators, whoever has decision-making rights? What needs to go right for us to be on that more productive path?

Mike: That's an extremely good question, and one that a huge number of people, including you, are wrestling with. It has a lot of dimensions. Part of it, I would say, is publicizing the opportunity set in a realistic but positive way. This relates to what some of us have chosen to call the diffusion challenge. I think you know what this is, but for your listeners: we're pretty sure these powerful tools are going to get picked up and used in advanced firms and technology, and probably in finance, or at least a subset of it. But there are very large employment sectors where there's at least a question, right? Government, pre-college education, maybe even college education, construction. I've looked at these sectors, and in previous rounds of digital adoption, they lagged. Healthcare is a very, very large one. Then there's traditional retail, hospitality, and so on-

Ben: And education.

Mike: And education. So the question is, is this all going to happen on autopilot or not? I think just posing the question is important.

Ben: Yeah.

Mike: What I've noticed is that there are three responses to this. One, on the American side, is that we're very reliant on market incentives and entrepreneurial activity. I don't mean to dismiss that; it's surely a component of the solution. But I suspect it's not the optimal strategy, and that complementary policy designed for information diffusion is needed, to make sure access isn't blocked by either a failure to understand or a lack of resources. To give a concrete example, I was very struck by the 15th Five-Year Plan in China. It's a very comprehensive document, and it's not a plan, as I think you know. It's indicative: this is where we think things should go and are going to go. A lot of expectations converge around these plans, so they affect investment behavior. On digital specifically, they clearly stated that they were not going to fall behind at the frontier, basically because of its importance for national security. But they also said they're going to make sure these tools are accessible and get adopted and used across the economy, and they named, I think, seven or eight sectors where they thought this ought to happen. I may not know how to do it, but once you set out on that course, there's some notion that the public sector may need to step in and do things that complement what the private sector does so well.

So that would be an example of working through these questions and asking: is there a role, and what is it? Asking serious questions: do people understand the technology and the opportunity set? Do they have access to the tools? If there are blockages, what are they? Is it excessive or dysfunctional regulation? There's probably no simple, one-sentence answer. But to the extent that we're engaged in a conversation about it, we're clear on the objective, right? We want the use to be benign, essentially augmenting opportunity sets for people, and we want it to broadly affect the economy.

The Apollo guys say something I think is important, and I've said it too: the diffusion issue is not separate from the question of whether we have an excessive investment boom going on. Whether the investment boom turns out to be excessive depends on the revenues that come in, and the revenues depend on how many people out there pick this up and use it on a sustainable basis, not fiddling around and then giving up. So globally, the macroeconomic productivity issue depends on the diffusion issue. The question of valuations and returns on investment depends on it. I view it as a core issue that requires the kind of conversation Erik Brynjolfsson and company called for in the letter they recently put together: technology people, policymakers, and economists contributing what we can. That kind of all-hands-on-deck approach is justified, I think. And I'll sign many letters.

Ben: Did you sign this one?

Mike: I signed it.

Ben: Okay. Cool. I don't want to get derailed by the letter.

Mike: No, no, no. It's very simple.

01:06:18 Advice for young economists

Ben: Thank you so much. This was so interesting, and thank you for your time. I have one final question for you. You've had such a varied career: research, academic administration, the growth commission, so many really interesting things. You've seen it all. Having seen what you've seen, do you have any advice for young economists looking to improve the world, or our understanding of it? What would you tell someone who is where you were 50-odd years ago?

Mike: Gosh. First of all, I would observe that the opportunity set for really impactful and really interesting research is very large right now. There are so many things we don't understand, where we need both empirical work, with imaginative use of data, and conceptual frameworks we might not have yet. I remember when the internet came along and it was pretty clear it was going to do something to the economy. People would ask me, are the models adequate? Is there enough structure in those models that this is just a parameter shift? And I said, probably not. The reason is that the art of model building in economics, as you well know, Ben, is getting rid of irrelevant stuff so you can see the important pieces working and interacting. If nothing changes for a long period of time, you can get rid of the parameters that characterize it. And if those parameters suddenly start to change, and I'm talking about deep informational structure, then you probably have to put them back in the models, which essentially means you need new models. Which is a long-winded way of saying I really think there's a huge amount of work to do.

What motivates most young academics, including most young economists, is that they really want to understand something we don't understand very well. It's what motivated me, and almost everybody I know. So the advice is: just go for it. I'm giving advice, but I don't think it's very important. Young people ask me a lot what they should do, what kind of training they should get. I say: in college, study whatever you get excited about. It doesn't matter, right? And broaden it. I remember a conversation with Derek Bok, a longtime president of Harvard, great guy, about where creativity comes from. I said, I don't know, but it looks like when you get things at the boundaries of fields that are supposed to be about the same subject but look different because the models are different, you get some kind of cognitive dissonance that leads to at least some creative activity. I think you see a lot of that in the world. So broaden it, but pursue what you love to do.

The only other thing I'd add: I remember that in graduate school we had a very talented group of people in the PhD class, as usual. We all went through the first two years and did the general exams, and everybody, I can't think of an exception, passed and did very well. But then we split. A bunch of people wrote PhD theses and went on to mostly academic careers, at least at the start, doing research. The other group went off to Wall Street, the World Bank, the IMF. And there was no difference between them in terms of smarts or capabilities that I was able to detect. It was just preferences. The people who went to the IMF, the World Bank, or Wall Street were contributing to an agenda that an organization had, which they presumably believed in, but the structure was there. Whereas an academic basically gets up in the morning and does whatever he or she feels like, and more or less doesn't care, in the short run, whether anybody else thinks it's interesting. I don't want to overstate that, but I think you understand the difference.

Those are deep differences in people's psychological characteristics. Some people are happy working away on something for a long time. The people who did the messenger RNA work, who got the Nobel Prize, were laboring away on their own for a long period, barely getting funding. But I'm sure that, apart from the anxiety of not being able to fund their research, they were having a great time, and it turned out to be stunningly important. Not everybody gets that lucky with their research, but I think it's about being honest with yourself on that particular question. Do you want to be part of an organization? Do you want to be something 20 years from now? Or are you happy just getting up in the morning and working on whatever you think is absolutely fascinating, once you find the thing that really motivates you? The happiest people I know professionally, including among economists, are people who literally love discovering things, and that's what motivates them.

Ben: Got to know thyself. That's great advice. Thank you so much for your time. This was so interesting, and thank you for being part of the show.

Mike: I enjoyed it thoroughly, Ben. Thanks for having me. Appreciate it.

Ben: Thank you.

Ben: The Economics of Work is brought to you by Revelio Labs, workforce data for research and benchmarking. If you enjoyed this episode, please remember to rate, review, and subscribe. It'll help us reach other curious listeners. I'm your host, Ben Zweig. Our producer is Cole Wagner. Thanks for listening.