Lawrence D.W. Schmidt | AI’s Impact on Labor and Wages

Julia Godinez:

It’s great to see you all. My name is Julia Godinez, and I’m an emerging technology risk specialist and member of the Federal Reserve Bank of San Francisco’s EmergingTech Economic Research Network working group. It is such a privilege to be with you all today. My work at the SF Fed focuses on emerging technologies and helping institutions understand and respond to technological change. So I’m delighted to open today’s session.

EERN was created to bring together academics, business economists, researchers, and the broader public to exchange ideas about how emerging technologies are reshaping our economy. Today, we’re focusing on one of the most discussed and timely questions, “how will AI alter the labor market and workers’ earnings?” We are honored to welcome Professor Larry Schmidt of MIT Sloan School of Management. However, as of today, he’s actually going to be faculty member of UCSD, University of California San Diego. And he will be joining us here in the 12th District.

Professor Schmidt’s work examines how technology influences labor demand, the distribution of earnings, and the risks faced by households and institutions. His research spans two major themes, the fundamental risk factors that affect the value of human capital and the consequences of imperfect risk sharing in labor and financial markets.

We will also discuss the underlying drivers of financial markets and an emphasis on decision-making, strategic complementarities, and information frictions. We’re thrilled to have him here today to discuss the impact of AI on labor and wages, including the distinction that happens to jobs and what happens to people.

Following Professor Schmidt’s presentation, we will both have live and pre-submitted questions in conversation with Huiyu Li, the co-head of EERN, and a research advisor here at the 12th District. As a reminder, today’s discussion will be live-streamed and recorded and will be available on our website afterwards.

As always, the views expressed today are those of the speakers and do not necessarily reflect the views of the Federal Reserve Bank of San Francisco or the Federal Reserve System. Now, let’s begin today’s session. Over to you, Huiyu.

Huiyu Li:

All right. There we go. It is my pleasure to welcome Larry, as I know him as Larry, to San Francisco Fed. He’s no stranger to the Bay Area. He has family here as well. I think I’ve always been very impressed with Larry’s work. He uses data that covers almost all workers and all firms in the economy to study important questions, like how innovation could affect workers and firms at the individual firm and worker level. And as you will see, there is an important distinction between what we see in the aggregate and what we see at the individual level. So without further ado, let me welcome Larry to the stage.

Lawrence D.W. Schmidt:

All right. Well, thank you so much for joining me this morning. I’m delighted to talk to you about this topic. And the title of the talk is AI’s Impact on Labor and Wages: Lessons From the Past and Thoughts for the Future. I do want to give a disclaimer that in some sense, we want to know about that question. We want to know what AI is going to do. Some of this depends on what you believe about science fiction and things like that. So I’m also going to bring in what I think I know about the past because it’s a lot easier to make predictions about the past than predictions about the future. And sorry for the bait and switch about … I’m now officially coming back to California, so delighted to be a little closer to all of you going forward.

So there’s another disclaimer, which is that I’m using various results which are coming from census data. They’ve all been disclosed, et cetera. It also involves a lot of joint work with my co-authors, so I don’t give them any blame for controversial things I might say today, but lots of credit for the research that went into it.

So if you’re like me, you’re kind of standing around at events that might look something like this. And people are walking around and looking at the impressive advances that we’ve seen with new AI technology. It seems like it can produce work product that looks a lot like what we’re doing. It also seems like maybe it could save a lot of costs. It also seems like it can do completely interesting but unnecessary things as well. And inevitably, at some point, this conversation goes to the doom and gloom direction, and we talk about these huge numbers, like 60% of people are exposed to AI. People are worried about getting replaced. There’s lots of financial stress that comes along with that.

And so in some sense, one of my goals today is to point out that this is a concern that’s been with us for millennia. Okay? So here’s a quote from Aristotle that, “If the shuttle would weave, the plectrum touch the lyre without a hand to guide them, chief workmen would not want servants, nor master slaves.” The Luddites, famously, were concerned about technology and displacement, et cetera. One thing that does feel different is who seems most worried, relative to maybe some of the waves of technology that have changed in the past. So there’s this new focus on AI, and particularly sort of how AI could really change white collar work, maybe blow up certain types of services, firms, et cetera. So that piece, for sure, seems new, even if the concern that technology can take away jobs is something that’s been with us for a long time.

I’d also say that there’s other perspectives on this. We’ve heard about tokenmaxxing. And Shiller made up another fake word, which I really like because I know exactly what he means: “This Doomsmaxxing Has Got to Stop.” That was his title of an op-ed in The New York Times. So I think one thing that’s useful is to just ask the question, “Is this fundamentally new? Is this time different? Or is this a version of something that’s been with us for a long period of time?” And so I want to tell you what I’ve learned about the data we do have to maybe inform that discussion.

So another thing I wanted to say is that regardless of what will happen. And I just want to be clear, the standard errors on any point forecast we’re going to make about what AI is going to do in 20 years are enormous, right? It could be … Exponential growth is a scary thing. So it could be a wonderful thing too, right? So we really don’t know. What we do know is that people are freaked out about it. So a recent Reuters survey said, “71% of US adults are concerned that AI could put too many people out of work permanently.” And then if you zoom in even further, because that’s just concern about the labor market broadly, you ask people about their own job. 18% of people think that it’s at least somewhat likely that their own job will be eliminated by AI or automation within five years.

I’d also say that there’s a whole bunch of surveys of firms out there. And for what it’s worth, the firm perspective is a lot less scary than the individual worker perspective. So most firms say it’s neutral. And there’s more firms that will say that it will actually lead to employment increases than employment decreases. Okay? That might also be a statement about the short run, not the long run, et cetera. But I think where are these fears coming from? And what can the data tell us about this? And so that’s what I’m going to try to get out of this talk.

And I really want to contrast two perspectives on how technology can influence the labor market. And the first one is what you typically hear and the discussion. We ask, “What will happen to jobs?” Right? And this is … If you want, if you think about how you’d measure this, you could measure surveys year after year. And measure, in repeated cross-sections, what people are doing and how many people are in each type of job title. What happens to the mix of jobs? And there, there’s a whole bunch of forces. And that’s going to be one of my objectives, is to at least use data on the first wave of AI. So everything up to ChatGPT, if you want, to understand these other forces. There are other forces that really can push in the other direction against the standard cocktail party narrative that we’re all going to die. Okay? Or at least economically, we’re going to die, even if we’ll live forever. I don’t know. That’s one possibility.

And actually, in data up to 2023, what’s interesting is that I find evidence consistent with that story that you hear at the cocktail party, that AI can take jobs, but there are these other forces. And it turns out that the same people who are exposed to AI-related displacement are also exposed to these reallocative forces that can actually push in the opposite direction. The other thing I want to emphasize is that this is not a new phenomenon. Okay? Technological displacement is something … I’ve used 150 years of census data. You can see it throughout that whole historical period. Okay? So this is something we’ve always been concerned about. And it’s always possible this time could be different, but it’s also maybe true that even if history doesn’t repeat, it rhymes. And so maybe these lessons from the past are likely to apply as well.

So that’s perspective number one, is what does this mean for jobs? And I would say I may be not as much of a doomsmaxxer as some people are when it comes to that, but I could be wrong. I just want to be clear. None of us really know. We’re going to have to wait and see because this is such a new technology.

But the second perspective is to say, “Well, look, jobs aren’t people,” right? If someone stops being a bank teller, they go on living and they do something else. What happens to that person? And in order to answer that question, you need different and better data. And frankly, we didn’t have that kind of data for a long time. And so we could only do the first thing and say, “What happens to jobs?” Jobs are not people.

And so I want to show you some facts from my work where we’ve been able to learn something about what someone’s job is. So what occupation are you involved in? And then we can follow those people over time, around the time that technology changes, and ask what happens to that person, because that’s what most of us actually care about, is what happens to ourselves, not to our job title. What happens to those people? And there, I’m going to look a little more like a doommaxxer. Okay? Not max, but I just want to point out that there’s a reason that people are freaked out, is that even if a job itself could be fine, technology often changes how those jobs … What they entail, the skills that you have to have, the tasks that you do.

And this, it turns out, can look like a risk factor from the perspective of the people who are doing those jobs before the technological change, right? If you’re really, really good at something and you’re getting paid really, really well for what you’re doing, change is not your friend, right? And that’s going to be the key insight there. So there’s this large, uninsurable, and concentrated downside risk. And it turns out that the folks who are most affected by this are also the folks who are most exposed to AI, which is highly educated, more experienced, better-paid individuals. Okay?

And so we see that, even in the last 40 years, so I’ll draw some parallels there. So one key takeaway from this is that a shock that can be good for a group overall can be very risky for the individuals who are part of that group before the shock. It’s just going to change what we need. It’s a force that redistributes resources across different workers in the labor market. And this is especially true for those people who the reason they’re getting well paid is because they’ve built up skills and expertise that are very valuable.

Okay. So let me first now talk a little bit about who is exposed to AI. Here, I think I’m standing on firmer ground because, again, this is a statement about the past, not a statement about the future. So this is from a paper with Menaka, Dimitris, and Bryan, where we basically use data on online resumes, so LinkedIn is going to be the big source of this information, in order to build a really detailed measure of how much firms are using AI and who’s exposed to those AI things. So we can talk about who’s exposed to AI.

How does this work? Lots of details in the paper. There are no details in my talk today. I’m sorry. Is basically we could go to, say, JPMorgan and we’ll find Ms. J. And Ms. J basically shows up because she’s using machine learning. It’s kind of tagged on her profile. And we’re going to do this using data on LinkedIn from 2014 to 2023. So one important caveat, all of this is before ChatGPT. Okay? So what happened before ChatGPT may not be the same as after. We can talk about that maybe in the Q&A, why ChatGPT may or may not be different. So this is pre-generative AI. Then we’re going to basically say, “Okay. What is Ms. J doing with AI?” And we’re going to use large language models (LLMs) to basically translate what she would talk about on her resume into what the firm is actually doing with AI. And LLMs turn out to be pretty good at that. Okay?

And then what we’re going to be able to do is ask, “Look at all the other people who work at JPMorgan. Most of them aren’t using AI directly. What happens to them?” And there, we’re going to basically build a measure of exposure. So what is exposure? It basically means that the things that this worker does in their jobs are similar to the things that the firm would be doing with this technology. Okay? I’m going to build a measure of exposure. That’s typically what people talk about in this literature. And that’s going to be based on how similar the text of what you do is to what O*NET, this database of job task description, says. Okay?

And so then basically, we’re going to have a list of how many people are exposed? How much AI adoption is the firm doing? And this is going to vary at the firm by occupation, by year level. And then in LinkedIn, I’m going to be able to basically follow these workers.And remember, the most exposed people to AI are going to tend to be white collar workers. These are the people who are on LinkedIn. Okay? So LinkedIn is a very good database for measuring that kind of thing.

All right. What do we find in the data? First is a stylized fact from the literature. So this is more of a replication than a new insight. Who’s using AI? It’s the biggest, best, most productive firms. Okay? So when you want to think about who’s doing this, it’s typically the market leaders. They have a lot of data and they have a lot of scale. And so it’s useful to do AI. So within any group you’ll look at, the bigger firms are going to do more AI than the rest. We can also ask, what are they doing with AI? And these are kind of … We basically split these descriptions of what the firms are doing into categories. And a lot of these look like what I would call intangible capital production of the firm. So running your firm better, doing better prediction, et cetera. Okay? These are the types of applications that we see the firms doing.

Okay? Next key stylized fact. Who is exposed to AI? Okay? So what I’m doing here is on the graph, we have, on the X-axis, a measure of the average pay of different groups of workers. And then we can ask what fraction of their tasks can AI do? And the key thing is that other than maybe at the very top, where a lot of those people are managers who are a difficult category to deal with, you basically see the higher paid you are, the more exposed to AI you are. As we’ll see, this looks different from technologies to the past. Okay? But who are the most exposed occupations? Again, these are people who are doing predictive stuff. These are hard jobs where you have a lot of specific training, et cetera. You work a lot with data and knowledge work.

And then the last thing that we do is we use the job postings of the firm. And those job postings of the firm basically allow us to ask the question of if AI can do a task, so the firm is investing in AI, do they stop asking for workers who have the relevant skillset to do that task? So we stop talking about certain types of programming, like working in Excel, if now you’ve moved everything into Python. And we find evidence for that, which is consistent with the idea that these sorts of cognitive tasks, many of which we thought machines couldn’t really do, now AI can do them and the machines can compete with workers. So it looks like there’s substitution, at least at the lowest level of aggregation, at the task level.

So then what we do is we build a big model. And if I was giving my normal version of the SF Fed talk, we’d spend half an hour on that model, but we’re not going to do that today. So it basically says there’s three forces that you want to estimate. And then it gives you a mapping from all this textual information that we have into measures of exposure. And the first one is the cocktail party story. Okay? This is basically if AI is good at the same tasks that workers are good at, you can offload those tasks away from workers on to machines. And that, all else constant, is going to mean that those workers … We need fewer of those workers around. Okay? So this is the task-level substitution force. And when people are building these measures of AI exposure, I think this is kind of what they have in mind. Fortunately, there’s two other boxes on the screen. Okay? So if it was just this, everyone in this room should be really worried because it’s your tasks that are most exposed to AI.

Now, force number one that pushes in the other direction. We do lots of things in our jobs, fortunately, not just one thing. And AI tends to be really good at the most painful things that we do in our jobs, so that’s also kind of nice for just happiness. But workers aren’t helpless. Workers can basically reallocate their time away from stuff that AI can do, like proofreading. I’m terrible at proofreading. AI is amazing at proofreading, right? Towards things that I like to do, like start ideas for new papers, et cetera. Okay? So there’s a measure of dispersion or concentration of the productivity gains from AI in a subset of tasks where as long as it’s the case that AI is only good at some things but not good at other things, you, as a worker, become more productive. That’s good. That makes you more useful to the firm. That’s something that kind of pushes in the other direction.

And then importantly, the same people who are exposed to AI are at firms that are leaders in adopting AI. Well, those firms, in our data set, look like they’re growing faster. And as a result of growing faster, making more stuff, they’re hiring more workers. And so it’s kind of this happy coincidence where the same people who you might think are most at risk for the downside of the substitution of competing with AI are also most exposed to the benefits from AI as well.

Okay? So how did this shake out? And again, this is not true job title by job title. There’s some job titles that are really hammered, but if you kind of … Remember I said that high-income people looked really scary? They were really exposed to AI. Low-income people, not so much. So if I basically just look at the first force, this direct exposure of your tasks to AI substitution, you see this strictly downward sloping profile, right? The people, the high-income workers, looks like there’d be many less of those jobs, many more low-income jobs. But again, I said there’s these two other things.

The first is those workers can re-optimize. They can reallocate their time towards things that AI is not good at. That makes them more productive. So it’s kind of like two steps backward, one step forward, where these reallocative benefits help those high-income workers and protect them partially from the shock. Next, the people who are exposed to AI are at firms that are doing a lot of investment in AI. And so it turns out those firms look like they’re growing relative to their competitors in the same industries. That’s the yellow line. You add these three things together, you go from the scary cocktail party story to a more nuanced picture. Okay? That was the key takeaway from that study.

So I think another important question that we need to ask when we’re thinking about the tone of this discussion related to AI is what does this really depend on? And a lot of it is going to come down to if you make a worker more productive, do you want more of them or less of them? And so my colleague, Roberto Rigobon, had this great example of toilet paper where he basically said, “Suppose that we got much more productive at producing toilet paper. Are we going to need more workers? Probably not. You just sort of need the same amount of toilet paper no matter what.” And so in that case, basically, demand for workers falls. But what fraction of the economy and what fraction of exposed tasks actually look like toilet paper? And so as long as we’re in a place where you make the worker more productive, you want more of them, these spillovers could potentially offset the direct substitution. And the point of our first set of estimates is that those spillovers looked big. They looked like they were big enough to, on average, offset those displacement losses.

The other thing that I wanted to mention is if you looked at those applications that I mentioned, a bunch of them relate to what I would call intangible capital production, right? So this is like running your firm better. And there’s some researchers who’ve emphasized the fact that we don’t invest as much in intangible capital as maybe we should because it’s hard to find the people you need to acquire that intangible capital. And so if AI unlocks that, surely that’s not toilet paper, right? If you could run a firm better and it was cheaper to run a firm more efficiently, you would want to do more of that. Okay?

A lot of stuff is in services, where it seems like there’s lots of scope to improve quality. And importantly, workers today can compete with workers in the future by helping to train AI algorithms, right? We can check the algorithm, create data for it, help it learn. That might be bad for workers 30 years from now, but those workers 30 years from now haven’t invested in skills, et cetera, so maybe we’re not so worried about them. Okay?

The other thing is just that we know that new technologies create new types of jobs. And it’s really easy to imagine the types of jobs that disappear. It’s harder to imagine the types of jobs that appear afterwards. Okay? So 60% of jobs in 2018 were from job titles that didn’t exist in 1940. 74% among white collar type professionals, you might think, are exposed to AI. Okay.

I’m going to speed up now, but I’m going to talk about some more … Zooming out a little bit, away from AI specifically towards what we’ve learned from the past. Okay? And so let me first zoom way, way out and use a century of patents. And basically, the point of this paper, which was in the Brookings last fall, was that we could use this same model I used to talk about AI to explain what happened to demand for different occupations over the last century. Okay?

So those same ideas, that this mean exposure is bad, but this dispersion across tasks is good, and spillovers offset the losses, this was true. It helped explain why certain types of occupations with more educated workers, you saw more of those in the past, fewer for less educated workers. It was also good for women. It was good for professional-services-type jobs. If you look at AI, that picture looks a little scarier. So this was our predictions about generative AI. And it looks like the red bars, or the AI predictions, they reverse a lot of what looked like happened in the last century. So from that perspective, generative AI looks a little scary. It looks like its technological change might have been skill bias for the last century. Generative AI, maybe less so.

Okay. Then there’s another distinction between different types of technologies that we emphasize in the paper with Leonid Kogan and Bryan Seegmiller and Dimitris Papanikolaou, where we basically said, “Look, there’s two types of technological change the literature has talked about.” One is automation-type technologies. These are ones where the technology is a very close substitute to what the worker is doing. Those are labor-saving technologies. And if you looked … We built a measure of this using textual analysis of patents. And there, you see this hump shape, which is basically that workers in the middle, especially around the 40th percentile, are most exposed to these labor-saving technological changes over the last 40 years. Okay?

And one example of this is the one-click buying patent from Amazon. So this is one of its most useful patents. And if you look at the number of people who were working as order clerks, which was the most exposed occupation, lots of very routine types of tasks, you saw that the demand for those people really fell off a cliff. Okay? So that’s an example of these automation-type technological changes, but there were other types of technological changes too. Think about Microsoft Excel, all sorts of software that allowed knowledge workers to do new things. This is what we called labor-augmenting technological change.

And here, we assumed, in line with lots of macro papers talking about skill-biased technological change, that this actually led to more demand for certain types of workers. So notice that the pattern of exposure there for these labor-augmenting technological change looks a lot like AI. So I’ll come back to that in just a second. But the point is if you looked at survey estimates, you look at repeated cross-section data, you see consistent with the labels that I attach to these things, innovations related to routine tasks that workers did were associated with declines in labor demand, whereas innovations related to non-routine tasks were associated with big increases in labor demand. So these were changes that were good for some workers. Okay?

And as I mentioned, this was my AI exposure picture. Notice that it’s not that different from this pattern of non-routine technological change. So it’s possible we can learn something from what was already happening for the last 40 years, and probably well before that, it’s just my data stopped before then, to these kinds of exposed workers. Okay? Now, AI might be more scary because AI also might be a substitute in these sorts of cognitive tasks, whereas before, it might have been more of a complement. Okay?

Now, in my last couple minutes, let me switch from jobs to people. Okay? And this is where we’re basically going to exploit the availability of these panel data sets where I can follow workers over time and see what happens to them. Okay? So these group-level effects that I just talked about might be ignoring an important redistribution that’s happening within those groups, or across peoples. Before, we were talking about workers competing with machines, but new technology also affects how workers compete with one another.

And in a world where workers can’t easily switch between jobs, typically if there’s changes in the demand, that’s going to show up from the individual worker’s perspective as a big shift in what they end up earning. Okay? And so because people enter and the composition of groups moves around, what happens to incumbents might actually be very different from what happens to the job overall. And so we’re going to look at workers over time in panel data and look at workers who are exposed to technological change. Okay. So one of the key estimates that we find there is this is a graph I already showed you on the left. This basically said that if you looked at this non-routine skill-biased-type technological change, that was good for the jobs. You saw more of those jobs, more dollars going into those jobs. But if you looked at the people who were doing those jobs before the technological change hit, those people are worse off.

For these routine-type technological changes, actually, the panel perspective and the repeated cross-section perspective are in line with one another. If a job just disappears, that sucks for the people who are involved, not surprisingly. Okay? Now, this effect is not uniform across workers. So if I looked at the workers who were doing routine-type tasks that were exposed to technological change, there, what I’m basically doing is splitting people into five buckets based on where they are in the lagged income distribution. There, those effects are kind of the same. If we just don’t need as much of something to be done in the labor market, that’s bad for the workers who are doing that, no matter whether they’re high-income or low-income workers. But if you look at these non-routine technological changes, that’s the blue bars, notice that the workers in the top 5% of the occupational income distribution are way harder hit. The same is true if you looked at college-educated workers versus not college-educated workers, or if you looked at older workers versus younger workers. Okay?

We see that these losses from these changes that are good for the group, like someone is going to be doing this job and someone is going to be making more money, but it’s not the person who was doing that job before. And one way you can see that is on the right, I’m showing a measure of job loss. So it’s basically you switch jobs, and you have a big earnings decline. And we see that that’s much more likely for these top workers in response to this non-routine technological change. Okay?

Now, there’s another reason why you might think that people feel differently about this from this perspective of what happens to jobs overall. So creative destruction has been around for a long time. And we typically talk about this in the context of firms, right? So we had railroads and canals. And then suddenly, the … So the railroads knocked out the ships. And then the cars and the trucks come in, and they knock out the value of the railroads, right? And there’s Uber versus taxi medallions, et cetera, right? All of these things mean that incumbent firms lose and some new entrants gain.

Well, what’s also true is that these shocks are really hard to hedge. So workers might be exposed to these shocks, as I’ll show you. Investors are exposed to these shocks. It’s not easy to get protection against stuff that doesn’t exist yet that might come in and blow up the stuff that you currently own. So these risks are very difficult to hedge. Okay? So what we do in this paper is we basically follow workers and we rely on an estimate from Kogan, Papanikolaou, Seru, and Stoffman, who basically showed that if I work in an industry, I’m a firm in an industry, and my competitors within the same industry develop a bunch of valuable new patents, this is bad for me. I’m going to grow less quickly, hire fewer people, et cetera.

We can ask … So that’s kind of what you’re seeing on the left. So in the top panel, this says, “If my firm gets a bunch of cool new patents, I’m going to hire more people, have faster profits, hire more workers.” But the bottom panel on the left says, “If my competitors within the same industry do the same thing, this is bad for me. They’re stealing market share from me.” Okay? And so that’s the result from their paper. But then what we do in this paper is say, “What happens to the workers?” Well, again, we’re going to look at the different workers within the firm. If you look at own firm innovation, both the gains and the losses are going to be particularly large for the highest income workers. These are the people who aren’t going to leave. They might be less happy or more happy, depending on what’s going on, but it’s really hard to get these people to re-optimize. They’ve got lots of human capital that’s tied to these specific firms.

And the other thing that you’ll notice, if you squint at this for a while, is the ratio of the blue bars on the left and the right is higher at the bottom relative to the top, which means that workers share, because share is not a good word when they’re talking about losses. They’re going to share more losses with firms. Whereas when firms do cool new things, one of the things the firms do is they hire lots of new employees. And those new employees dilute the existing employees. This is especially true for the most novel patents that the firms do that look very distinct from past patents. Okay? This is a graph I won’t have time to talk about, but it basically says workers would be willing to pay money … They look like Luddites in the model. They would be willing to pay money to not have cool new technologies emerge because they’re so worried about the displacement risk.

And this relates to a discussion that we’ve heard in financial markets about AI stocks looking expensive. Are we in a bubble? Maybe. Certainly possible. I saw only AI-related billboards on my drive to the Fed, so could be. Another possibility is that people might be overpaying for AI stocks, and knowing this, because they’re really worried about AI-related displacement. And so you want to … If you’re worried that this thing’s going to blow you up, you want to own it as protection in that instance. And this is connected with the literature in finance on the value premium that shows that that’s linked with displacement risk. Okay? So a common theme is basically if you think about human capital, which is the most valuable asset that people have, human capital, just like physical capital, is exposed to the risk of creative destruction. People have firm-specific skills, they have vintage-specific skills. And technology changes the mix of skills that are going to be needed, and this creates winners and losers. And so in incomplete markets, and I don’t know about you guys, but I find it very difficult to buy insurance against income shocks that I experience, then technology looks like a risk factor for individual workers’ perspective, even if it’s good in the aggregate. Okay?

And related to this, actually, measures of income risk, especially for high-skilled workers, are rising. And this is a force that might also help us understand why interest rates have been declining over the last 20, 30 years, because basically, the only way you can insure against your human capital to protect yourself from displacement is to save. And so if people are really worried about idiosyncratic risk, downside risk, they’re going to potentially bid up prices of assets, and that’s going to pull interest rates down. Okay? So I’m done.

If you’ll give me one minute, I wanted to make one statement for business folks because there’s this discussion that people are having. And lots of people are asking executives, “What are you going to do with AI?” And I would just say maybe let’s change the tone a little bit. I understand why the people who sell AI products want to talk about capturing 15% of the labor share or whatever and offloading that to AI. But for everyone else, you’re scaring your workers. And one of the points that we made was that it might be useful to think about the gains that come from AI. You’re making your workers all more productive. That’s usually a good thing. So think about ways of augmenting what your workers do. Don’t just replace them. And think about expanding production possibilities. Can you build better products, better services, make better decisions? Do more with this cool new technology, rather than to do the same with less? Because in our model, it’s reallocation of your time that generates these productivity gains.

And related to that, firms have some choice about how to insure workers. One of the roles of the firm is to provide insurance to workers against shocks. And so it might be that if you want AI to work better, it could be useful to provide some degree of downside protection to your workers so that it’s like, “Look, yes, you can train AI, and I see that that’s valuable to me, but I’m also not going to have you train AI and then immediately fire you the next day.” So think about ways of managing that displacement, even within the firm.

All right. So I’ll leave these up here. But thank you very much.

Huiyu Li:

All right. Thank you so much. Thank you, Larry, for such a rich discussion. As I said, I think it’s really helpful to see what’s happening at the individual and worker level because in the aggregate, we don’t see very strong decline in labor share compared to the pace of technological progress. And also, aggregate employment has been rising despite all the technological changes, but we see it in your work that individuals are facing a lot of risk.

So let me start the discussion by just understanding better what you have said, that there were a lot of information there, and also to get to know you a little bit better. You are an economist, so can you give us some examples of, in your work, what type of tasks you have, like routine, non-routine, and how they could be substituted by or complemented by technology?

Lawrence D.W. Schmidt:

Yeah. So I guess if you … Well, I’m an economist. I’m also an economist who’s just moving right now. So one of the things that you see is there’s massive amounts of paperwork that you have to deal with as a part of important transactions in your daily life. And there’s many ways in which even IT technologies of the past massively reduce the cost of doing these things. You think about Docusign and just how much time that’s saved, and no one had to deal with the paper, et cetera. And surely, once we sort out the legal complexities of all of this, AI is going to be able to further reduce those kinds of pain points. So I think about that as fairly routine. I would also think about copy editing, checking things, downloading data and just getting it read in, all those kinds of things. For me, often when I code, the hardest thing for me is installing the stupid packages. Thus far, I’ve relied on RAs who are younger than me and better at that, and that’s great. And now, I have AI which can maybe do that installation part.

So for sure, there’s all these things. And, honestly, what I love about AI is that despite the fact that it’s a little scary how good it is at some things, the things that it’s the best at are the things that I’m the worst at or that I get most tired easily when doing them. And so I’ve found that it’s made my research life a lot more pleasant. So there’s the routine stuff, like copy editing, for sure. It’s been amazing. Yeah.

Huiyu Li:

Yes. Maybe this is a good segue to talk a little bit about GenAI versus AI. I don’t know if you’ve been experimenting with GenAI. Do you see GenAI having different effects on the types of tasks that you do or complement it differently? Or maybe there’s some tasks that used to be complementary to AI, but now with GenAI, GenAI could actually replace you in doing those tasks?

Lawrence D.W. Schmidt:

Yeah. So in the model we wrote down, that in the data we use, we were really thinking about predictive analytics, machine learning, things like that. That was the original wave. Typically, you need a big mountain of data. And then if you have a big mountain of data, you can make it very useful for some specific types of tasks. And what is definitely different about generative AI is the scope for what it can be helpful with right out of the box. And so in our model, what was helpful for workers is the fact that AI is very rarely good at all of the things that a worker does. And so there’s places where people can go to re-optimize their time. And so if you thought about the managers, high-skilled folks who are exposed to predictive analytics and stuff, great. They spend less time fiddling with their predictive model, and that lets them just spend more time making decisions, et cetera.

What’s a little scarier about generative AI is the scope of the tasks that it can potentially affect. In our model, what that might look at, as basically saying that those concentration effects, those reallocative benefits might be lower with generative AI. At the same time, there is this wonderful thing that happens, which is that if you’re exposed to adoption of AI, yes, that means that you might be competing with a machine that’s quite good at something, but that also typically means that you just became more productive and your firm just became more productive, right? And so we tend to omit that part from the discussion. And so if generative AI is just better, then those spillovers also should be bigger, in principle. And I think a lot of the challenge is routing the tasks to these things. It’s not even necessarily completing the tasks. Yes, it might be better at completing them than we are, but knowing what you want to do next and why, and all those kinds of things, I still see a lot of benefits for humans in the loop there.

So, we’re working on measures. And you definitely see … The same types of people typically are exposed to both, generative and not, but there are some differences, like legal, for example, very low exposure in our earlier measure, whereas legal seems like it’s potentially very exposed to GenAI.

Huiyu Li:

Yeah. I guess in our profession, coming up with ideas is very important. And if AI can do everything else, and we just need to sit there and come up with ideas, it makes our ideas more valuable, so then maybe our earnings won’t be hit as much, hopefully.

Lawrence D.W. Schmidt:

Hopefully. Yeah. And hopefully, if we become more productive, the cost of producing research goes down, then we’ll demand more research. Right?

Huiyu Li:

Hopefully.

Lawrence D.W. Schmidt:

It’s not toilet paper. Yeah. So exactly.

Huiyu Li:

All right. I think there’s a lot of discussion about what will happen to net employment right now. We’ve seen past technologies improving, progressing, but net employment hasn’t declined. And I think maybe it’s related to entrants actually benefit. People entering to labor market benefiting, even if incumbents suffer earning loss. Do you think something might be different for GenAI? Because there are some studies showing that maybe workers entering the labor market right now is not doing very well.

Lawrence D.W. Schmidt:

Yeah. So the entry-level question is really fascinating. And there’s a bunch of forces in place. So one thing that I think we often forget, this is the other side of research that I do, but something else happened around the same time that ChatGPT happened, which is that interest rates, real interest rates went up dramatically. And there’s a lot of uncertainty. And the macroeconomic outlook has sort of been weird. And there’s a lot of caution that I see. And something that I’ve shown in other work with Maarten Meeuwis, Dimitris Papanikolaou, and Jon Rothbaum is basically that those are exactly the types of workers who you would expect to get hit when interest rates are high. Because typically, new people, you’re paying some cost upfront in order to train them and get them up to speed. You realize the benefits in the future. The benefits of that match are actually very sensitive to interest rates.

And so I would think part of why that … And you actually see that the decline happens a little before ChatGPT. And so that might be part of the story. I’m not saying it’s the whole story. And then typically, the new group has the biggest incentives to come in and invest and learn how to use the technology. And so one thing that’s weird about AI, or generative AI in particular, is that if you know how to do something, it sort of levers your expertise, and you can check it. And that aspect of things maybe protects incumbents in a way that might not have quite been true in the past. It also could be that the tasks that entry-level people do are simpler and more reliably copied by generative AI. So you might think that the exposure could be a little higher of the tasks that new people do, and the concentration might be lower. So those are forces that, in our model, would say, “Okay. Maybe we need fewer of these workers.”

At the same time, who are the AI natives? It’s all of these new people. And so which force is going to win? This is complicated. It’s hard to figure out. And I think maybe the role of the firm should be to find ways, bring in these new people who get how the tools work but maybe don’t have the relevant expertise, and partner them with the senior people and get that leverage working properly. And so my guess is a lot of this has to do with the outlook. That’s not true across the board with all the jobs, but I think in the long run, this is something where if it looks like other technological changes, I’m willing to bet that younger people are going to end up doing a great job of deploying AI, probably better than the folks who’ve been in the labor market for a while.

Huiyu Li:

Yeah. That’s a really great outlook for the younger folks. I do feel that because I think now, we’ve been more experienced in the research. And I use research assistants. I’m sure you do as well. And I feel like that for GenAI, many of the things I used to ask research assistant do, I don’t need to anymore because GenAI does it, but it doesn’t mean I don’t need them because they are quicker at actually learning GenAI and deploying GenAI. So they’re just doing a different type of task. Rather than replacing RAs overall, they’re just doing something different.

Lawrence D.W. Schmidt:

Exactly. And if you think about, a lot of these productivity gains are coming from figuring out how to set up these loops, right? The self-reinforcing cycles of how to learn from past mistakes, that’s complicated. I haven’t fully figured that out. I might be faster at checking ChatGPT’s derivation than my RAs are, but that’s only a part of the job. And I’ve had some wonderful experiences working with RAs who themselves have learned how to use ChatGPT effectively or other GenAI tools in order to do model extensions and code up new things and build lots of visualizations. And I must say, an RA who’s good at AI is much better than just me trying to battle it out with ChatGPT or Claude.

So again, I think one stylized fact that’s useful for us to keep in mind about technological changes in the past is that things take forever, right? We have these periods of time where for 50 years, some people are still using horses on farms, even though tractors have existed. And how can that happen? I think it’s because of some of this stuff, that it’s costly to adopt these things. You don’t want to break things that are working. And so the change tends to be gradual. And yes, AI feels like it’s all going faster. But I think to get it to really work to the degree of accuracy where you trust it, it’s still hard and you still need data and you still need people to check it, et cetera. And so my hope is that that same slowness will happen here, but …

Huiyu Li:

Yes. I’m sure, especially for larger enterprises, for them to really fully embrace GenAI, there’s so many things that they need to check, both on the legal side, well, just to make sure it works.

Lawrence D.W. Schmidt:

Exactly. That’s something that’s very new about AI, is knowing whether it works or not, is … And often, it looks really close, but then something important is wrong. And that piece of it, it does make it feel very distinct from some of the past technologies.

Huiyu Li:

The other part of your talk that I really appreciate, as you were discussing this, is that we usually talk about income inequality. And people say that high-income people earn so much more than low-income people, but you also show that they’re actually facing also more risk. So you could be high-income one day and then lose most of your earnings if your firm gets displaced by a competitor or if a lot of your skills gets destroyed by GenAI. So I guess, what do you think … This is speculative. How do you think GenAI may affect the income distribution and also the risks that people face?

Lawrence D.W. Schmidt:

Yes. That’s the million-dollar question. So again, let’s start with where we’re on more solid ground, which is if this looks like the past. If this looks like the past, then I think about it as the higher up you are, even within a job title, you think about doctor, you give a general practitioner around the median and then you’ll have the ultra-specialized person who’s treating … They’re the absolute best at some specific kind of surgery at the top. And so typically, what happens is that that person at the top is … They’re earning some sort of very high compensation because their skills are really scarce. And so change is very scary for that person because even a small change of your rank in the productivity distribution can be associated with a massive change in your earnings. And what we’ve seen, this is related to work with Kyle Herkenhoff, Carter Braxton, and Jon Rothbaum, is basically those types of people, and this tends to be associated with cognitive types of tasks and the types of jobs where people are using a lot of computers in the past, those people’s risks have just continued to go up.

So the income inequality has gone up, but those people are also moving around. And so if you were to follow that person’s earnings over time, and particularly the persistent component, the thing that is not just about fluctuations this year versus next year, but the next 10 years or the rest of my career’s worth of earnings, that risk has gone up dramatically. And it’s linked with technological change in the sense that it’s the same places where the frontier is moving faster that those people’s risks have gone up.

That was actually peaked around the IT Revolution. It was declining a bit in the last 10 years. I’m willing to forecast that AI is going to take my line graphs and increase their risk again dramatically because almost surely, there’s going to be people who had incredibly valuable expertise. And that expertise might have just been subsumed by AI. And that’s going to create some losers, just like all of the other technological changes in the past. So I don’t see why we wouldn’t expect displacement.

It is maybe somewhat comforting that the people who are most exposed to that displacement risk are the people who have some degree of risk-bearing capacity typically to begin with. So one of the other points we made is that these people are saving more. And this is a force that we think is maybe pushing natural interest rates down.

So I see no reason why those forces wouldn’t persist, but then there’s these interesting things about generative AI that could go either way. So on the one hand, it seems like AI can be a substitute for expertise in some contexts. So you can take people who are very experienced, AI doesn’t help them so much. Then you can take people who have never done something, AI seems to improve the quality of their work product. That seems like a leveling-the-playing-field-type force, but that’s in a world where you force everyone to use AI with the same intensity and in the same way. It does feel like it’s experience enhancing, so that seems like a force which could easily blow up within occupation inequality.

So I think the true answer is we don’t know. It’s probably going to be heterogeneous. It’s going to be different in some industries and occupations relative to others, but I’m willing to bet that there’s going to be … Regardless of what happens with the aggregate, and I hope I tried to offset a little bit of that negativity that you hear at the cocktail parties and in the popular press, but the aggregate might be fine. And there’s still going to be a lot of people who are really hurt by this. And that’s been true with technological change for a long time. And maybe, on the other hand, the fact that this has been that way for a century should make us a little less afraid because people’s incomes have always been volatile, right? 10% of people in a given year are going to take a 40% pay cut. We don’t think about that very often, but it’s a part of life. And I think technology is one of those reasons why a person’s earnings can decline dramatically because they could have been really good at something, and suddenly we don’t need that skillset anymore.

So maybe the fact that it’s always been that way means that, yes, it’s scary, but that’s a risk we’ve had to stare in the face already. And maybe the enemy that we have some familiarity with is better than the one that we don’t know at all.

Huiyu Li:

Yes, for sure. Yeah. I think part of GenAI being scary, just somehow it feels unfamiliar. But making it more familiar, maybe we can think about how to deal with it. Yes.

Lawrence D.W. Schmidt:

Yeah. To be clear, all the science fiction scenarios, we’ve all read different books like that. It could happen. That is scary. And that’s totally different. But if it looks like the pattern of diffusion and adoption from the past, it’s typically slow and it’s typically going to be those best paid people who are going to be most exposed. And they’re still going to be able to eat and all those kinds of things, right? These are big losses, but it’s not like they’re going to be utterly catastrophic.

Huiyu Li:

Yeah. Not doom.

Lawrence D.W. Schmidt:

Yeah. You might change your standard of living a little bit, but you’re not necessarily going from the C-suite to living on the streets or things like that.

Huiyu Li:

Yes. Okay. Let me switch gear to some questions. We actually had a lot of questions, so we probably won’t cover all of them. But let me just first go to the audience that we have here with us today. So one question is, “There was an article in The Times today about a tech support nonprofit that will work with community college systems and others to retrain workers displaced. Do you have any impression, improvement, or ideas in this area?”

Lawrence D.W. Schmidt:

This is going to be the million-dollar question, right? Where are people going to go? What I would say is that it’s encouraging that the people who are most exposed to AI are also the people who have been in jobs where that’s always been a part of it. You’ve always had to know how to deal with the changing responsibilities of your jobs. Some people say, “Oh. Now we don’t need education anymore because ChatGPT can just teach us all instead.” And I’m not so sure about that. Absolutely, we’re going to need education as an important part of it, and especially if it’s those tasks that you used to be able to learn on the job because they were things that needed to get done and you could learn quickly.

If that rung of the ladder is gone now, people are going to have to come in knowing how to do these things as is. And so it seems like that’s going to be a role for education. At the same time, it’s hard. I’m talking with many of my friends who are educators. And a lot of the tools we had to get students to do work, especially more complicated project-type things, people can just plug it in into AI. And so we have no way of knowing whether they’ve really learned how to do that or not.

Huiyu Li:

That’s true.

Lawrence D.W. Schmidt:

I think we’re going to solve that problem. And I’m guessing AI is going to be a part of that solution in the sense that AI could be a part of your student group and make things more interactive and fun and all that, but it’s going to take time. So we’re going to have to figure that out because people are going to need to redeploy their skills. But I think it’s … I take some solace in the fact that usually, the demo is amazing, right? With AI. I want to remodel a kitchen. And so sometimes you see these examples of people have AI do these remodels, and it looks beautiful. And then I tried to do it for the actual one at the house that I’m looking at. And it looked a little weird and it wasn’t that useful, right? The gap between the demo and the full-scale product is pretty large still. And we’re going to need people to fix all of those little things.

I also think people can identify the pain points in their job because often AI can be used to make the jobs more pleasant. And so people are only going to be willing to do that if they don’t feel like the immediate response of the firm is going to be, “Cool. Well, now we don’t need you anymore,” right? So we have to manage this carefully. But I think there’s really a huge opportunity here. And I would like our discussion to be a little more centered on, “What are all of the amazing opportunities to do cool, great, new things and to have workers be way more productive and happy?” rather than just, “Oh. How can we take exactly what we were doing before and just do it with fewer people?”

Huiyu Li:

Thank you. We actually get a lot of audience online that are students. And also, I see many RAs, interns, actually, here as well. So a common question we get is do you have any advice for people trying to accumulate skills? What should they focus on to insure themselves against job loss in the future from GenAI?

Lawrence D.W. Schmidt:

So I think the key in all interactions with AI is to use it as a way of enhancing your thought process. It’s not a substitute for thinking. And the more that the goal is to just offload a task to AI, (A) it’s less successful anyway, it just doesn’t work, but (B), then you’re not making yourself more informed and useful in that process. And so I think having AI teach you things quickly is an amazing use of it. AI as a tutor is unbelievable. It can even write problem set questions and help you solve them and check them. But it is not a substitute for learning how to do it yourself. In some sense, the value of knowing how to do it yourself might be even higher in a world with AI adoption being high because if you can check it, then you can suddenly scale it, right? But if you can’t, you’re useless, right?

And so I think that this idea of, “Oh. I shouldn’t major in something because AI can do it all.” It’s like, no, that was never the point. People majored in economics, and they do wonderful things in the business world. I don’t know how often they draw supply and demand graphs on the whiteboard. It’s not that important for their day to day. But the thinking, the framework that came from having done that a couple times in college was useful. And then also, in college, one of the main things you learn how to do is to learn itself, right? And so people need to learn how to be really good at learning. And I think a lot of the high-paid jobs in the world have always involved constantly learning. And so basically, the key thing is to use generative AI to supercharge your ability to learn new things and do new things.

And so yeah, absolutely go and take the classes, learn, do the problem sets, woodshed on these things because we as humans can’t just download the new weights. We have to sort it out ourselves. So go and do that and enjoy it. And maybe these tools can help you learn it in different ways, which at least for me, I needed to hear something explained three different ways. ChatGPT can explain it 30 different ways now. And so that capacity is amazing, but I definitely wouldn’t want to have it do my problem set for me.

Huiyu Li:

Thank you very much. That’s all the time we have for today. Thank you so much for coming, both online and in person. We have more events coming up, so please stay tuned and sign up for … Subscribe to EERN from our website. Thank you very much.