← The AI Policy Podcast
The AI Policy Podcast
The Impact of AI on Labor with Harry Holzer
01:07:10
Economist Harry Holzer joins to discuss how AI is set to transform labor.
Loading summary
Transcript90 lines
- [00:00]
A
Foreign.
- [00:11]
B
Welcome back to the AI Policy Podcast. Today we've got a big episode on everything at the intersection of AI jobs and AI's impact on labor. And we have a especially qualified guest to talk about this. He is the former chief economist of the U.S. department of labor and also a professor at Georgetown University University's McCourt School of Public Policy. I'm speaking about Harry Holzer, who is the co author of a paper recently published on AI's impact of labor. But he has a long and distinguished career studying the labor market from a bunch of different angles, including technology's impact on it. And so that's why we're so glad to have him. Harry, thanks for coming on the AI Policy Podcast.
- [00:56]
A
Good to be here, Greg.
- [00:57]
B
Thank you. So I talked with you, I talked with the audience just a moment ago and explaining your bio about how you've had an enormously distinguished career. But I want to ask, you've been a labor economist for more than 40 years. What were some of the most important experiences that shaped your interests and where you are today? And I'm especially keen if any of those intersect with the technological evolution over the past four decades.
- [01:24]
A
Well, I'll throw out two One from earlier in my life, I actually grew up in a rural area on a chicken farm, the old fashioned kind with long chicken coops in the backyards and where all the chicken were free range.
- [01:41]
B
This is going to come as a shock to some of the listeners, but I grew up in Kansas and we also had chickens, but they were just like the chickens from my kindergarten class that we decided to take over. And we weren't actually farmers.
- [01:52]
A
Right. And my parents are immigrants, they're Holocaust survivors. So I had a sense that life can be hard for people, sometimes economically, sometimes socially or emotionally. And that led to my interest in a career where I can maybe help people adjust economically. I would say the other thing, once I actually got my PhD and started doing work, I've always focused on fairly disadvantaged groups in the job market, low wage workers, people of color, did a lot of work on young black men, especially black men without college degrees. And you get a sense of, you know, it's a complicated issue. People make a lot of choices, but you get a sense that a lot of people have very, very limited opportunities. And if we can help them, and including for reasons like automation, among other things, so if we could help them a little bit, that that would be a good thing as well. And that has also motivated my work.
- [02:47]
B
Yeah. And I think, you know, today when we talk About AI automation. Oftentimes folks are thinking at like the high end of the labor market, like automating programmers. But for most of the past 100 years, automation and even the disruptions from trade has been primarily targeted at the low wage side of the labor market.
- [03:06]
A
Right. We have a term for that in labor economics. We call it skill bias. And most of the automation historically has been skill biased in favor of more skilled workers, whatever that is, at any point in time. Certainly the digital revolution, which started about 45 years ago, seems to have really benefited college graduates and professionals, but really replaced a lot of people without college. People working on assembly lines, people working in offices. And of course the big question is, will AI follow that pattern or not?
- [03:41]
B
Yeah. And so before we dive into the AI impacts, I kind of want to give folks, I want you to give folks rather, if you can put on your professor hat, Labor Economics 101 and Labor Policy 101. So let's start with like the economics side of the equation. I think most people have applied for a job themselves and, you know, see the headline unemployment numbers. But like, what are the fundamentals of the labor market? How does it work?
- [04:09]
A
Well, so, so there's a big interesting question that a very prominent, prominent labor economist named David Autor, he titled one of his papers about a decade ago, why are there still so many jobs? In other words, we've had so many waves of automation, and ever since the Luddites, some people have predicted that all the jobs would disappear. Why? Why does that always turn out to be wrong? Historically, the answer is there's a couple things going on in the economy and the labor market. First of all, automation makes the workplace more productive, and it makes the workers who can use the automation more productive. But when productivity goes up, that almost always means costs go down, prices for consumers go down. As long as the markets are competitive, that means the income, the price, adjusted incomes or wealth of consumers goes up. They spend more money. Sometimes they spend money on the same products that they've been buying. Now they just buy a lot more of it. Sometimes they spend the money on new products, things like smartphones that didn't exist 25 years ago. And so that means there's a lot more demand. And sometimes. So the increase in demand and in spending by consumers creates more jobs for a lot of workers. And sometimes, sometimes, if you're talking about a product that's relatively new, it's very expensive. Early on, that was true of automobiles. More than a century ago, it was true of personal computers and smartphones. Thirty Years ago, smartphones didn't even exist. But then the drop in price can be so dramatic that all of a sudden it goes from being a luxury good to something that can be sold on the mass market. And that's when jobs really grow very rapidly. It was true when Henry Ford created the assembly line. And all of a sudden, for the first time, cars were affordable to the middle class, to working class, et cetera. It's been true of PCs and smartphones in more recent decades. So some jobs disappear. Sometimes new job categories of jobs open up. So Henry Ford's assembly line put the horse and buggy operators and the. And the craftsmen building, you know, wagons to hitch to horses. It made all that obsolete, but created hundreds of thousands of new jobs for auto workers that didn't exist before. And I can give you many historical examples like that.
- [06:40]
B
Very cool.
- [06:41]
A
That's the economics.
- [06:42]
B
Yeah. And I think we're going to come back to a lot of those terms of art you introduced, you know, not just supply and demand, but elasticity of supply, elasticity of demand, a lot of these things. But now I want to get the second, you know, part, because you already touched on technology a good amount. But what about on the policy side of the equation? So I think everybody knows there is a Department of labor, but I don't think everybody knows what it is in doing all day, how many people are working there. And then what's the role of the chief economist, the position that you occupied?
- [07:13]
A
Well, the Department of Labor oversees federal programs and federal policies for workers for labor. The biggest bureau there is the Employment and Training Administration. They administer a lot of the job training programs that the Department of Labor oversees. They oversee unemployment insurance, apprenticeship programs. And then of course, there's many other. There's programs like occupational safety and health, mine safety, many other, all of which touch on some kind of policy. And that's very important because when, when automation comes along, some of the skills people have become obsolete. And I didn't talk yet about compliments and substitutes. Some people already have the skills to work with the technology, their compliments. Some people are replaced, their substitutes. But you want more workers to adjust by picking up new skills that make them compliments and compliments that enable them to use and benefit from the new automation. Workers need help doing that. They can do some of that on their own. But that's why we have all kinds of job training programs, career and technical education, and things like unemployment insurance for people who lose their jobs for no fault of their own and need some time to adjust to that. So all of that is in the purview of the Department of Labor.
- [08:38]
B
Yeah. So if I could. I mean, I think there's three functions that I heard from you. If I could play back to you what I think I heard first, the Department of Labor has a critical role at the federal level of enforcing federal labor regulations. And that could be something like osha, which you mentioned, Occupational Health and Safety. Or it could be something like civil rights related, for example, could be civil.
- [09:05]
A
Rights or things like wage and hour rules, minimum wage.
- [09:08]
B
Are you actually letting your people have time off? Are you actually doing. Right, so there's a regulatory aspect, then there's the intervention aspect where the US Government is actually trying to shape the nature of the labor market, which could be by these retraining programs that you described. It could be through unemployment insurance, trying to make unemployment a less painful financial disaster. And then I think the third one is trying to study and understand the labor market, which is a data collection and publication of statistics and advising Congress, advising the executive branch kind of thing. Are those the three sort of main functions of the Department of Labor. Did I get it right?
- [09:51]
A
Yeah, I think that's a good summary. Yeah.
- [09:52]
B
Yeah. Okay, cool. And just out of curiosity, like, how many humans are at the Department of Labor? Do you remember what it was when you were there or what it is?
- [10:00]
A
I don't, I don't remember. But it's important to remember that most of the humans aren't in Washington D.C. in the DOO building. They're spread around the country.
- [10:09]
B
Right. If you're an OSHA inspector, you need to be where the factories are.
- [10:11]
A
Exactly, exactly. Or collecting data or administering programs like unemployment insurance. So it's, it's bigger than the one building in Washington.
- [10:22]
B
Got it. Okay, so now let's shift gears and talk about AI's impact on the labor market. And I want to start with a report that came out on October 1st from the Yale Budget Lab and Brookings, where you are a non resident senior fellow. And the report was titled Evaluating the Impact of AI on the Labor Market Current State of Affairs. And so this is very much tied to, you know, ChatGPT being out for about three years right now looking at the generative AI revolution. And their conclusion was that at least the metrics that they studied indicate that the broader labor market has not experienced a discernible disruption since ChatGPT's release 33 months ago, undercutting fears that AI automation is currently. That's a key word, eroding the demand for cognitive labor across the economy. So the authors looked at changes in, like, the economy's occupational mix to evaluate the impact of AI. Could you explain to our audience what is the occupational mix and like, why is it a useful metric in thinking about this?
- [11:27]
A
Well, the issue of who is exposed to AI on the job, whether it's in a positive way or a negative way, happens at the occupational level. At the job level, every occupation involves doing a set of tasks, and it's the tasks that automation might take over. And in this case, AI might end up doing a lot of the, you know, whether it's writing early drafts of certain things or composing, composing articles or music or anything like that. So we tend to look at, at the occupational level. Now, there's an important caveat to this. The Yale Lab paper, which is a good paper, looked at very quite broad occupational categories. And I wondered if the categories are a little too broad to really be able to capture what's going on. A different paper, a paper by Eric Brynjolfsson of Stanford and some of his co authors, looked much more narrowly. They looked at young college grads, people in their mid-20s, roughly, and in some very narrowly defined occupations like customer service or software development, which he picked because he thought those were the occupations that would be affected. And there does seem to be some softening of the job market for those people. But again, you had to cut the data very narrowly by demographic group and by occupation to find some effects. And we're not even sure if the effects he found were really caused by AI or something else, like something to.
- [13:03]
B
Happen at the same time as a lot of AI.
- [13:05]
A
That's right. I mean, we know, we know the job market has softened the last year or two.
- [13:08]
B
Yeah. Not least of which in the tech sector, where you had this massive hiring spree associated with COVID and all the demands of a world that wasn't going outside. And so everything digital was in high demand.
- [13:21]
A
That's right.
- [13:21]
B
And the hiring move associated with that, it's. It's difficult to look at the data and know for sure that the impact you're looking at is AI driven versus other exogenous factors. I mean, that said, there are anecdotes, there are some noteworthy CEOs who are talking about big, you know, AI impacts. Just to give you one example, Amazon CEO Andy Jassy wrote an email in June to all Amazon employees saying, quote, we expect AI. And there, I'm adding AI. He said he was saying it, but we expect AI will reduce our total corporate workforce as we get efficiency gains from using AI. Extensively across the company. And then just the other day this week, Fortune reported that, quote, Amazon is preparing to cut as much as 15% of its human resources staff, with additional layoffs likely in other divisions. Now, they weren't specifically attributing that to AI. You know, there's no official statement on that, but I thought it was noteworthy that they really do think that their workforce is going to be slimming down as AI goes. So if the Yale paper is saying that there's no obvious broad based impacts yet, and the Stanford paper is saying maybe there are some impacts in these, you know, categories where we think it's the most exposed to automation risk, you sort of look at the expectations of executives like leadership in Amazon and there's plenty of others. I mean, there's plenty of startups that are saying, like, you know, we're going to hire 10 employees and that's going to be the company for the rest of the time. You know, as big as we get, we're going to stick with 10. And so I'm curious, you know, what you make of all this. I mean, you were talking about the Brynjolfsson paper. What is your current sense of where we are right now in the automation story and where we might be headed over the next few years?
- [15:09]
A
So I think it's very early on, you know, maybe where the Digital Revolution and PCs were in the early 1980s and employers haven't yet figured out the best way to implement this. And even the software developers, I think are still waiting for where's the market going to go on this and how should they develop this? And an important part of the story is that employers and the AI developers are going to have a lot of discretion in terms of how they mold the product and how they implement it. And it can be done in a more worker friendly fashion, where it's designed for workers to complement it or in a more substitution or what they call automation as opposed to augmentation fashion, we don't know yet. You know, they'll be responding to the market and of course the other issue, and I'm sure there will be Amazon's one company, there might be lots of other, and it's a big company, lots of other companies where you might have a slimming down role. But on the other hand, number one, if they invest in at least some retraining of the workers, the workers might be able to pick up new tasks that they don't do right now. So if in the Amazon warehouse, drones are now flying around distributing or robots, but there might be Some other product where demand goes way up that could be useful. And are these workers, do employers create new jobs and then do they retrain their existing workforce for some of those new tasks and some of those new functions? I just think it's too early to tell on all of that.
- [16:50]
B
Yeah. And I think one thing is the variables in the equation of how much of an impact does AI have on the economy do not exclusively depend on how good is AI at this specific moment. Right. Like the ceiling for how much impact AI can have is set by the capabilities of AI. But the floor and the present state, a lot of other variables can impact that. And so let's just talk about some, like, prior technological revolutions, including, like, revolutions that you've, you know, observed over the course of your career and had to. Had the opportunity to study in the labor market. So there's. There's one example that a mentor of mine, you know, clerked for the Supreme Court and was a lawyer at a prestigious law firm. And he said when he got there, he got his own computer. He got, like an IBM, you know, a computer, and he did his own typing. And it was like a scandal among the partners of the law firm that he didn't have a secretary take dictation. He was like, you're. Are you. You're not doing your own typing, are you? And the point here being that, like, the technology can be ready, but the workforce has to also be ready, the culture has to be ready. The businesses have to want to adopt it, and those kinds of things can be slow, even if the actual potential of the technology is to be fast. And I think, to give another example that you mentioned, the computer revolution of the 1980s in factory manufacturing, one of the places where that showed up. And, you know, I'm a business school graduate, so I think about it in a lot of these terms, but one of the places where that showed up was in enterprise resource planning software. So if you've ever heard of the European software giant SAP, they're a big player in this industry, and they help you look at, you know, what's all the inventory in my factory? What are we planning on manufacturing? What are we planning on buying from our suppliers? How many workers do we have right now, and what could we make and at what price could we charge? So it's sort of helping you as a factory manager decide what orders to accept. And historically, this was done on, like, massive paper spreadsheets, you know, with the accounting department and the supply chain department, et cetera, working together to plan all of this. Stuff and it took forever. And so they only did it once a month. And then once you had computers logging all of this and tagging it to data that was updated more frequently, you know, that went from a process that takes a large department a month to calculate to now the computer can just do it in an hour. And what's so interesting is that the companies that did this were like, well, why aren't we getting these massive productivity benefits? And it's because, number one, like did you get rid of the department or did you keep all those people so you didn't get any wage savings? And then number two, are you still only calculating it once a month? You know, if it's, if all you have to do is just press go on the computer, then you can run that calculation every day and have a lot more, you know, strategically informed decision making process as you're deciding, you know, what contracts to accept and at what price to accept them. And so the point here is that there was this whole process re engineering that had to take place for the benefits of the automation technology to be seen. And so those are examples of just big sources of friction in the adoption of the technology and therefore friction in the impact of that technology on the labor market. So I guess my question to you is, you know, you were in the seat and were a prominent labor economist during the, well, you were kind of late to the computer revolution, but to the Internet revolution, to the mobile revolution, to the cloud based software revolution, and to plenty of other, I'm sure, technological disruptions outside of just digital technology. So, you know, what are, what are some salient examples in your own mind of how this story plays out and how we should think about friction and other factors.
- [21:00]
A
Well, I think everything you've said is accurate. I want to quote a very famous quote from a wonderful economist named Robert Solow, who was a Nobel laureate, whom I knew quite well. And he was looking at the productivity numbers in the economy in the 80s which were not very good. And he said, but people are talking a lot about these productivity effects. He says, why do we hear and see productivity gains from computers everywhere except in the numbers, the economic numbers. But in the 90s the numbers changed and you did start to see it, but it took a decade to work out all of the things you're working out for employers to figure out how to best use these machines, which workers to let go, which workers to retrain, all those things. So that takes time and by the way, then all kinds of new products and new startups for people to spend Their money on that create jobs that don't even exist today. But you know, another area much more recently where you've seen something like that is talk about autonomous vehicles. Five, seven years ago, before the pandemic, people were assuming that, that, that human driven cars would be a thing of the past. Very quickly.
- [22:14]
B
Oh yeah, I mean I can give you two fun quotes on that. Number one, I remember very quickly, very clearly because I made a bet related to it. In the summer of 2016, the CEO of Ford said that you will have, Ford will be selling cars where you can get in, sit down, fall asleep, be driven to your destination. I think he said, in five years. So by 2021. And then you know that that same summer Elon Musk predicted that full self driving was only one year away. And he has made that same prediction every year in the subsequent 10 years that autonomous vehicles. Now I actually see, I see Waymos, you know, on my drive to work pretty frequently at this, at this point here in Washington D.C. so autonomous driving is here, but it's, it didn't take, it wasn't nearly as fast as many folks expected.
- [23:01]
A
That's right. And there's a whole, what you might broadly call an infrastructure, some of those infrastructure. Part of that infrastructure is trained workers, you know, workers who know how to service and maintain the Waymo vehicles. Lots of other things, adapting, you know, the light system and all kinds of safety protocols and things like that. And that just takes a lot more time. So we might, it might still come to pass in the relatively near future. But all that highlights the point you made about frictions and adjustment time.
- [23:34]
B
So we've been talking about mostly through the lens of unemployment and layoffs as a potential impact of AI. And I want to get your reaction to one of the most extreme predictions that's out there on this. And this comes from Anthropic CEO Dario Amadai. So in May he told Axios that he believed that AI could eliminate half of all entry level white collar jobs within five years, causing unemployment to spike between 10 and 20%. By contrast, Nvidia CEO Jensen Huang said that AI will change jobs but not replace humans. So can you just, you know, sort of walk us through these, these various theories and how AI could lead to mass unemployment. Why people like Jensen Huang and I believe yourself are skeptical of the sort of mass unemployment outcome.
- [24:30]
A
Well, first of all, we've heard the song before, going all the way back to the Luddites 200 years ago who were convinced that automation and some of them had actually been displaced by automation.
- [24:43]
B
Yeah, no, they really did lose their jobs.
- [24:45]
A
That was a real thing.
- [24:47]
B
Yeah, the Luddite revolution. For those who aren't, who don't remember from high school history, this was a group of, I believe they were weavers. Like they took, they took wool and they spun it into fabric. And that had previously been a cottage industry where you do it at home. But the Industrial revolution, and this is even the pre steam engine industrial revolution, this is just the sort of like building machines that do stuff at scale and having a factory system put them all in business. Either said like you, you come to the cities and you work in the factories or you don't have a job anymore. And so they decided to stage a little riot where they smashed the machines. And that's like the, you know, in D.C. policy debates on technology and unemployment. It's sort of a rule that somebody, you know, has to invoke the Luddite story at some point.
- [25:32]
A
That's right. That's right. So. But there have been much more recent examples of people, including economists, predicting big waves of joblessness. I mean, 20 years ago, a lot of people were saying offshoring of all work to China is going to wipe out millions and millions of jobs. Didn't happen. So I'd say I'm a little skeptical of Dario's story, you know, for two reasons. Number one, there's a difference between, in theory, a machine can do a certain task versus the quality of the work, of implementing it in a specific workplace to a specific set of tasks that workers do. And the human pieces of that might still be needed to check the accuracy of the work, to check how it's represented. So I think, think, I think there's a big leap, you know, like as there was with autonomous vehicles where people were predicting. So that's one reason I'm skeptical of Dario. But of course it also, it also avoids the whole adjustment process of workers retraining, getting new skills, employers investing and retraining them. So, so that's why I'm, I'm skeptical that he's going to be right about that. And I might be a little closer to, to the gentleman from Nvidia now. Now then there's a question. Is this time different? Is there something about AI that qualitatively makes it very different from all those waves of automation over two, two and a half centuries? And if there's something to worry about, it's the depth of what AI can do and the breadth of tasks and the speed of it, the fact that it's constantly Reinventing itself and constantly. So in the past a lot of the, like if you got laid off from an assembly line and a lot of those workers never did come back. But at least in theory, you could learn one new task and go out in the job market. The worry with AI is that every year so you can make an adjustment, learn a new task, and then AI might take that over a year or two away. So there is a possibility that AI is qualitatively different and that the disruption in the labor market will not only be greater but worse for workers. That'll be like a treadmill and they're always running to try to catch up and never, never quite catch up. We don't know. I remain more optimistic than Dario, but on the other hand, I don't think we should discount the possibility of a lot of disruption, a lot of displacement and workers might need a lot of assistance to adapt. And that's an interesting thing about, you know, people debated whether AI will be worse for college grads or like the digital revolution, which was worse for non college grads. Young college grads might face more exposure to AI, might face more potential substitution from AI. On the other hand, my guess is that they will be better at adapting to whatever out there, getting new skills, learning new ways. And by the way, I think virtual training will make new skill acquisition easier and more accessible. But that'll be where I think college grads have the leg up. They will adapt better even if they face more exposure on their jobs.
- [28:53]
B
So there's a lot in what you just said and I kind of want to disentangle some different pieces of it. The first is timing, right? Dario is saying within five years, right? So he's saying this is coming on fast and it's going to have a big impact. The second thing is, you know, he's talking about unemployment spiking. And I want to distinguish here between AI short term job impacts and really just automations short term impacts and automations long term impacts. So it is definitely the case, right, that automation has led to job losses. That is obviously true. Many, many, many, many times. However, automation leading to unemployment in the aggregate economy sense is kind of a qualitatively different phenomenon. So saying like hey, you 500 workers whose current job is to, you know, lift stuff off a ship and take a bag off of ship and then put it in the bag on the truck. Well, when we have containers and automated, you know, automated loaders, like that is not a job anymore. Robots do that job, period. But it also dramatically increases the productivity of shipping and receiving across the entire economy. And so those job losses don't necessarily mean macroeconomic unemployment. In fact, they probably lead to economic growth that creates additional demand for labor.
- [30:30]
A
That's right, yeah.
- [30:31]
B
The second thing I think is interesting here is you've talked about the distinction between, you know, augmentation and automation and I, it's, it's not obvious to me that there is a difference. So I kind of want you to explain to me what the difference is because, or, sorry, it's not obvious to me that there's a meaningful difference in terms of labor market impacts between automation and augmentation. So attractor is like obviously to me at least a case of augmentation. One guy swinging a scythe, you know, can harvest so or, or plowing, you know, with a hand pushed plow, can plow, you know, so many acres per day. And then you give this man a tractor with a plow on it and suddenly he can, you know, plow 100 times that much acreage. And so he is being augmented. His individual productivity has increased a lot. But, but if instead of him being augmented, it was just the case that there was this tractor that drove itself right alongside him and he wasn't allowed to touch it, but it could increase the amount of plowing per day. It seems to me that the labor market impact would be almost identical and that what determines the within sector job impacts are not augmentation versus automation type phenomenon. They have more to do with the elasticity of demand in the end market. Right. We can eat twice as much food as people a hundred years ago, but we can't eat a hundred times as much food as people a hundred years ago. By contrast, think about the production of video. This podcast, which we're doing as a video podcast. In the 1970s, producing this product would have required probably at least 40 people. We have three producers, and only one of them is involved in the technical side of the equation. And so the point being that the productivity for video production has gone up massively, but consumption of video has gone up massively and prices have gone down massively. And so in one case, in the farm case, the number of Americans that are employed on farms has gone down a lot since a hundred years ago. I mean, I think it's gone down by more than tenfold Even though the amount of food that Americans grow and produce has increased by a huge, huge amount. So this is a case of like increased productivity did destroy those jobs in that sector, but in the economy it was fine because those agricultural workers became manufacturing workers or service workers. And then in the video case, it was fine because consumption went up by so much that productivity didn't. The increased productivity actually means that there are, while there's very few people working as a cameraman today, that's just because the reporter, you know, does their own cameraman work and a producer now wears 10 hats instead of one because the productivity is so high with all the technological enablers. So help us untangle these different phenomenon in the context of AI, both where we are today and where we might be if we continue seeing additional technological progress.
- [33:54]
A
Well, there's a lot there.
- [33:56]
B
I threw a lot at you.
- [33:57]
A
Yeah, you know, I think there will be all kinds of different examples. So for instance, if we go back to the early introduction of personal computers or a couple decades later, smartphones, those barely existed, and when they existed, they were very expensive. So one answer to your question is if the prices come way down, where is the product in its life cycle? Is it, if it's relatively early and way too expensive for anyone to use it much, then the elasticity of demand might be quite high because if the price comes down enough for middle and working class people to afford it, the market could explode. And of course, and that's what happened also with Henry Ford's right model.
- [34:48]
B
The original cars were for aristocrats exclusively and Henry Ford made it a middle class phenomenon.
- [34:54]
A
Right. So. So that's a case where the product existed, but it was way too expensive for most people to afford. And, and so a lot of new product demand and then worker demand was created when the price goes down. A different category. You know, so smartphones are a little different because you had flip phones and things like that before, before that people had blackberries. But the smartphone was really a very new product.
- [35:18]
B
The deviation, it was a qualitative increase in performance and user experience.
- [35:21]
A
Yeah, a whole different animal really. So that's a new product that didn't exist, you know, replacing all the flip phones and the blackberries and all those things. But the new product comes along and we're going to see a lot of that with AI, I believe, all kinds of all kinds of new products, new services that don't even exist now, and a lot of new jobs. There might be other cases where on the job, you know, automation will eliminate AI, will eliminate some part of what people do. But then the role of workers to oversee the product. I mean, right now there's enormous accuracy issues with almost everything that AI writes. I'm guessing that might be true for a while or whether it's relevant in a particular context. So that kind of works. So then, you know, maybe some of the, the people right now who are doing the lower end writing will do more of the editing, more of the oversight, things like that. And of course, in that case, it'll be very important whether the employer in those establishments invests in retraining them. And frankly, you know, American employers have always been a little more skeptical of investing in their workers and it's gone down over time. There's a lot of different reasons for that. You know, I think employers have a bias towards investing in training their professional and managerial employees, maybe their technical workers, and much less for their non college, either blue collar or pink collar workers. But employers are going to have to make judgments, okay, who is, who does it make sense to reinvest in and who not. You know, the good thing for a lot of workers is that if the employer has been happy with their performance, you know, it's costly to a firm to go out and recruit and hire and retrain somebody. So, so there might be a reason to keep a lot of their incumbent workers and to invest in them. On the other hand, in many ways the tax system through depreciation of new investment, tilts the playing field in favor of automation against incumbent workers. So that works the other way. So these are all the variables in the equations that you described, all the different ways in which new jobs will be developed. Sometimes in the same industry, sometimes in the same firms, sometimes in different ones that don't even exist now. And it's just a belief that all of that will relieve mass unemployment. If we go back to the unemployment issue a little bit, this is what economists have often called structural unemployment. When you had people doing well, a full employment economy. And then a big shock comes that automates the workplace and it makes a lot of workers obsolete at one time. That certainly happened in the digital revolution, sort of the combination of the automation, the robotics and what we call the China shock really made several million manufacturing workers obsolete. And they were often concentrated in company towns and specific regions, say of the Midwest.
- [38:29]
B
Yeah, like, like, you know, obviously the, the Rust belt didn't used to be called the rust belt.
- [38:33]
A
That's right.
- [38:34]
B
You know, called like the manufacturing center of America.
- [38:36]
A
That's right. So if you looked at, at specific cities like Detroit or Cleveland or specific states, you did see a bump up in unemployment that could be called structural. Most economists will tell you that historically a lot of economists overstate structural unemployment at the aggregate level, at the level of the whole country. Some people will be displaced and something, you know, if you're a displaced worker above, say, the age of 40, the odds of being retrained for something new do go down. So there might be all these new jobs being created that they might not be the people getting. And there's been an important new paper by economists David Autor and Gordon Hansen. You know, they introduced the term the China shock and they looked at those labor markets and then many years later they went back and said those labor markets have recovered, but the people originally displaced in many cases were not the ones that, that got the new jobs. Did we have mass unemployment? Was there a big jump in structural unemployment that lasted for any in the aggregate? The answer was no. If you looked at certain specific regions and specific industries, you see a little more evidence. So I'm guessing it will be more like that. Now, AI is not going to be as concentrated in a particular purpose technology. That's right. It's a general purpose technology. It'll be spread much more. In some ways that's good thing, right? Because it means that, you know, it's not like you're going to take out Detroit and the Detroit economy is going to have so much trouble. It's spread more widely and maybe that means the adjustment processes will be a little easier.
- [40:14]
B
Yeah. So I'm, I'm, my sort of current hypothesis is that, you know, I was talking before about how technological progress is not the only thing that matters, but I do think it is a thing that matters and matters quite a bit, especially when you, you know, come to scenario planning for some of the extremes that basically the CEO of every AI company is predicting right now. Now obviously they have an incentive financially to get people excited about what's just a year or two away when you're trying to raise hundreds of billions of dollars to fund all this capex. But just to be blunt, right, there's plenty of folks out there like Elon Musk who are saying that we're a handful of years away from AI systems that are as smart as the smartest human in every category of intelligence. So let's just, let's, we're going to talk about that, but let's bracket it for one second and just say I believe that because of all the sort of friction items that we talked about before, that even if all technological progress in AI were to magically stop today and we were sort of frozen at the current level of technology, we are still nowhere near, you know, the overall economic, we are nowhere near the potential economic impacts of AI because every industry sort of has to figure out what works and what doesn't work in adopting this technology and which shortcomings can be accommodated and, you know, addressed, you know, remediated in meaningful ways and which ones sort of mean AI is not a good fit for your industry. I think there's like many years of just sort of digesting the current state of the technology. And my prediction is that that would lead to, to job loss and it would lead to job creation, and the net effect would probably be no additional increase in employment in the aggregate sense of that term. But now let's think about the more extreme scenarios that some folks are talking about, like Elon Musk, like Dari Amadai, like Sam Altman, the CEO of OpenAI. And the thought experiment that comes to mind for me is, you know, coming back again to the transition point between farms before tractors and farms after tractors. And I think there's some really interesting data points here. So in 1900, 41% of the US labor force worked in agriculture. In 1970, 4% of the workforce in the United States worked in agriculture. So a huge drop off in people working on farms. Now, that was fine. And the reason why it was fine, I mean, it was not fun for the farmers in question. I'm sure it was a very painful disruption. But in the macroeconomic sense, it was fine because America got much richer over that period. And unemployment, you know, was not meaningfully differentiated between those two times. We were able to have very low unemployment and growing average incomes in, in the 1970s, 70s. However, you know, those human workers, they retrained to go work in manufacturing where their dexterous hands were very useful, and they retrained to go work in the service sector where their brains were very useful. So maybe they weren't competitive with the tractor at pushing a plow, but they, you know, were way better than what was available in terms of automation in the manufacturing and service sector. However, there was a different labor force in the United States that had neither dexterous hands nor big brains, and that was the population of horses and mules. And so between 1900 and 1960, I couldn't find the number for 1970, but the number of horses and mules that were employed in labor in 1900 was 21 million, and in 1960 it was 3 million. So they lost their jobs and they didn't get new ones. Right. And so the point here basically being that there was, there was no new job to be created for them because they were not competitive in any way with anything. And so I want you to just like, play out the thought experiment here that Elon Musk is asking us all to consider when he says superintelligence is around the corner. And it doesn't have to necessarily, you don't even have to agree with him that it's around the corner. Let's just say it's going to happen sometime in the next 100 years. What does you know, your economic theory and your body of experience tell you? How does this story change as AI's capabilities get more and more extreme, far beyond where we are today, and you start approaching or surpassing human level intelligence in every category of intelligence?
- [45:15]
A
Well, I love the example of the horses and mules versus humans. And lucky for us, humans, we are different from the horses and mules along a couple of key dimensions. Number one, there's a whole lot more dimensions to what we can do with our hands, with our brains. You know, horses and mules could only do one thing, pretty much. And once you replace that, and of course, they didn't have the ability to adapt and humans have the ability, you can start off being a substitute, then, then you can learn some new skills, new tasks, and become a complement. Now, it might take some real adjustment, you know, in your story about agriculture versus manufacturing, first of all, a lot of those people had to move to cities, streets, from the countryside. That alone took a lot of time. And, you know, there's always this question of, I'll go back to what I said before about when an area recovers from a big shock, is it the people who are displaced who get the new jobs? It might be their kids, it might be only their kids, if those kids went to college, etc. So some people take a lasting people.
- [46:24]
B
Who moved to that area.
- [46:25]
A
Right? That's right. That's right. So, you know, some people take a lasting hit. Hit. Other people, I think, will figure out how to adapt. I think the younger you are, the better chance you have of adapting. Maybe having a strong basic level of skills also means you can adapt to a lot of different tasks and things. But then let me address the thing about, you know, if the machines are smarter than us on every single dimension and historically, every bit of automation. Like, like I remember when calculators, you know, when I was in high school and calculators came out and again, way.
- [46:58]
B
Better at math than me.
- [46:59]
A
They were way, you know, and math teachers, who is wrestling with, do I still need to teach some of this stuff if you can press a button on a calculator? So it was smart on those particular dimensions, but many, many other Dimensions, it was not smarter than us. So if Elon Musk is saying, well, the machines are going to be smarter than us on almost everything, I'm skeptical of that. And you know, I think that there will remain something about humanness and human centered abilities and tasks that will remain qualitatively different from a machine for a long time. I think first of all, in terms of exercising judgment, looking at all of the relevant context when a judgment is made, I think AI over time will be able to look at more and more of those factors. But will it have the judgment to know how much wait to put on something? And maybe this context is different than maybe, maybe in my town, it's a different context than in a, a different location or a different city or state. You know, there's so many factors. I think humans will just be better at that. Anything involved in human emotion. And, and I know, I know that AI, you know, people are talking about people dating their, their AI machines and.
- [48:16]
B
You know, having them as their therapists and their emotional support companions.
- [48:20]
A
You know, I just gotta believe that at the end of the day there is something about humanness that will distinguish it and continue to distinguish it. And even the knowledge that it might be a machine talking to you in one case, maybe even a machine that can do sex in some sense, but the knowledge that it's not a real person on the other end, I think will change the human experience and we'll continue to give human beings a leg of.
- [48:49]
B
So I think, I think there's, there's, if I could disentangle your argument, there's, there's two essential claims that you're making. One is that, I mean, basically you're saying humans have advantages over machines that are likely to endure. And I'm oversimplifying here, but you can, you can put those into two categories, arbitrary advantages and actual advantages. So what I mean by arbitrary advantages is Magnus Carlson is not the best chess player on earth. He is the best human chess player on earth. He, he cannot beat his own phone, right, at chess, ever. He would lose a thousand to zero. But his phone does not have an awesome, you know, ability to attract, to fill stadiums and, you know, millions of YouTube viewers to watch, you know, the phone play chess. There is only a market for watching humans play chess. There is essentially no market for watching AI play. Now that is an arbitrary preference, but it is clearly a real one, right? And so throughout the, throughout the economy, this type of phenomenon might show up again and again and again in terms of customer support representatives or you Know, do I want to get my massage from a robo massage thing or do I want to have a human or do I want to have an AI therapist or do I want to have a human therapist? Blah blah, blah, blah. You can imagine that this arbitrary preference for humanness will be an enduring economic advantage with meaningful outcomes. And then the second thing is the not non arbitrary advantages. So you're asserting that machines will be fundamentally incapable not just next year, but maybe even a hundred years from now. They will continue to be inferior to humans in delivering emotional tied services basically. Or their, their overall judgment of weaving together all aspects of their intelligence. Your assertive that even 100 years from now, no matter how much technological progress we make, the sort of 2 kg of gray matter inside my skull will be a superior intelligence device. Do I understand you correctly?
- [50:57]
A
Yes, I think that's right. Now can I prove that? No. It's what my gut is telling me. And it might be, it might be that a big chunk of what workers do today the machines will be better at and that essential humanness that I think makes us unique. Again, it might be involved in all kinds of new job categories that don't exist today. But there's so many examples. You mentioned the sports example. I'm sort of a fanatical follower of the Los Angeles Dodgers even though I grew up in New Jersey because Sandy Koufax was my hero as a kid. Now one can imagine a set of robots outperforming the Dodgers or all these other teams. Am I going to get excited to watch them? Of course not.
- [51:42]
B
Yeah, I might watch that once. Right. I'm not going to watch that three times a week.
- [51:46]
A
Exactly. Exactly. So you know, I can imagine some current jobs that I just can't imagine a machine really replacing a person and creating the same demand. Or maybe all kinds of new functions that will people will come up with as new products get created and as new things are needed. And I think it's going to be some combination of that that will mean that humans will not become completely obsolete in the long run as well as the short run.
- [52:14]
B
Yeah. So I'll just state that I disagree with you that in the long term that humans are going to have some kind of enduring advantage. Right. I think a hundred years from now, unless we get awesome brain implants that make us smarter, I would expect that the computers are better at us in everything except for the arbitrary preferences where I certainly agree with you. Right. That people might just like a robot is not eligible to apply for the job of my wife, my wife has already taken that job. And I'm not taking applications right, for a new job, no matter how compelling the robot's pitch is. Okay, so now let's shift gears and talk about some of the policy parts of this story. So you were one of 40 leading economists who signed a letter to the current labor secretary, Chavez D. Remmer, calling for better data on AI's impact. So we've been speculating a lot here, and part of the reason why we're speculating is the data is inadequate. And the letter says, quote, we are concerned the government is not adequately prepared for the collection of high quality economic data that will inform policy to address the workforce issues AI creates. So what data is the government currently collecting and where do these efforts fall short, you know, in your understanding in terms of the impact on AI of.
- [53:34]
A
AI, rather most of the labor market data collected? Well, there, there's different kinds of data. There's surveys and then there's administrative data that already exist. So, for instance, a lot of the survey data come from the Bureau of Labor Statistics in the Department of Labor or from the department.
- [53:52]
B
Great agency that does phenomenal work. I just want to say I think.
- [53:55]
A
It is, it's, it's, its reputation has been unfairly sullied in the last bunch of days, but that's where a lot of the survey data come from. And then there are these administrative data. Like every quarter, employers, if their workers are covered by unemployment insurance, they have to fill out data and send it to their state agencies for unemployment insurance coverage. And that's become a very important source of data. So there's survey data, there's administrative data, and of course, the third data is stuff that gets scraped from the Internet. You know, big data of, of different kinds of, you know, I think all of them need to adjust. All of them have some potential, but all of them will need adjustment. The surveys will simply have to ask all kinds of questions that they don't ask now about what do you do in your job and are you using some program with AI in the work you do? I think the administrative data, a lot of states are linking their quarterly earnings data to their state education data. So you might be able to, you might be able to follow a worker who studied AI in college, look at the firm that employs them later, and then what happens to those workers versus other workers who have worked at that firm previously. So you can see, you know, who.
- [55:17]
B
That's something the government does not do now, but you, you think would give policymakers information they should want and do need. If they could sort of start tracking this because of thing.
- [55:27]
A
Yeah, the capability is there and of course, the new data. There's all kinds of new ways of collecting these data that will come up and, and you know, and already in the private sector, you know, we have companies that collect job vacancy statistics, that collect people's resumes. One can imagine public private partnerships that make those data more useful to people like me, policy, you know, researchers and policy analysts. So. So we definitely need new data that don't exist. Some of them you can exist, you can tweak existing sources of data, and in other cases, maybe all kinds of new sources will come along that we're not even aware of. But we need it. We need all those things.
- [56:08]
B
I find that completely persuasive. So you. We're coming up on time here, and I want to be respectful of your time, but you also recently co authored a white paper titled Proactively Developing and Assisting the Workforce in the Age of AI, which included a range of policy recommendations. And I should mention, you know, you had a lot of co authors on this piece, but can you just give us an overview of the paper and some of its main recommendations?
- [56:34]
A
Well, the paper, you know, we had about 15 people, so people bought a lot of different expertise that no one of us would have had. And first of all, the paper just reviewed the whole issue of AI capabilities. You know, what's the best that AI can do now, now, five years from now, 10 years from now? What is the evidence to date? Number one on productivity growth, and number two on employment and wages and disemployment. And again, recognizing that it's very, very early to make any of those judgments. And then we had, you know, what are the relevant policies? The section that I had, you know, one of the, one of the, one of the sections was on what we just talked about new data and new research. I had a lot of input into the section on education and training. You know, that's my interest and expertise area. So I broadly think of sort of three big buckets. One is kind of basic K through 12 education, higher education. It's like the foundation that everybody gets before they enter the labor market. The second bucket is stuff that happens on the job. So if AI takes over, someone's task asks, will the employer retrain that worker for a new task, will they let them go? And then, of course, the third bucket is if people do get displaced or if they think a displacement is coming, all kinds of adaptations that will help them retrain. Maybe at a community college, something like lifelong learning accounts, maybe automation adjustment systems modeled after a program called Trade Adjustment Assistance. All the things that people will need if in fact they see displacement coming or they get displaced. So in each of those areas, you know, the report talks about things government can do to make those kinds of education and training more complementary with the AI out of the starting blocks to nudge, incentivize, assist employers to take their current incumbent workers seriously. You know, maybe, like, maybe you could subsidize retraining, you could even tax worker displacement. Although that didn't make it into the report because it's more controversial and the evidence base is more limited. There's all kinds of ways in which government can do each of those things. Government can provide guidelines, and a lot of this happens at the state level, you know, for their K through 12 systems, their career in tech ed and their high schools, their public, public community colleges and universities, guidelines for how they should be adjusting the curricula in response to AI. And of course, a lot of that's going to involve interactions between the education institutions and the business community to get a sense of what's coming down the road and how to prepare people for that. So there's things you can do in all three of those buckets. There's a whole area a lot of people right now are talking about. If in fact it's true that young workers, young college grads, are going to have trouble getting entry level jobs and gaining that work experience, maybe we really need to make a push around work based learning, internships, apprenticeships, maybe they will play a more important role for giving workers that early workspace.
- [59:51]
B
The thesis being right now that entry level jobs, you know, basically allow you to do a certain category of work and you're partly compensated in dollars and you're partly compensated in terms of training. And if the AI replaces that entry level work, then nobody is giving you that training and nobody's giving you those dollars. And so you might need a new system that incentivizes companies or other organizations to provide the training that can help workers make the jump from university to, well, whatever the new entry level will be, which is not the current sort of state of what entry level is.
- [60:32]
A
That's exactly right. I mean, labor economists have known for quite a while that that sort of first five to ten years in the job market right after one complete schooling is a very important period for all the reasons that you just said. You know, you learn all kinds of new tasks and you develop a track record by which employers can judge you. And, and many Times if you really want to make progress, you don't stay in the same job, you shift to the same job in a different company, maybe to a different job. And if AI makes all of that harder for a young graduate to obtain, it would be good to have an alternative mechanism like apprenticeships or internships, maybe to fill that.
- [61:12]
B
So I want to ask about like the, the, specifically the support for displaced workers because I feel like these programs kind of have a bad reputation in policy circles. Right. Because it was one of the great theses for why the China shock would not be so shocking. I think in the 1990s, part of the sort of argument to liberalize trade regimes was saying, yes, there will be displaced workers. Yes, the textile workers of North Carolina are going to be devastated, but the federal government is going to dump a bunch of money into North Carolina to retrain these workers. Workers to work in some other industry that is not textile manufacturing, which is going to go to low wage countries. And there's nothing really to be done about it. And you know, you're a labor economist, you know a hundred times more about this than I do, and this is even your area of expertise within labor economics. But, but my sense is that there's a kind of generalized dissatisfaction with the average amount of impact that these retraining programs have for displaced workers compared to, you know, know, the marketing slogans of the policy in the 1990s. So, you know, as, as you think about it, like, are there any salient examples where workforce displacement programs and retraining programs, like, knocked it out of the park? What does this look like when it goes right.
- [62:34]
A
Well, it's a tougher job in a lot of ways than say, a training program for disadvantaged young people. You're talking about, about a lot of people now who might be in their 40s, their 50s, much hard to retrain at that point. Where they live might be a depressed region. The new jobs, they might have to move somewhere else at that age, people are more sedentary.
- [62:59]
B
Yeah. You have kids.
- [63:01]
A
That's right. There's a lot of great.
- [63:03]
B
When you're that age.
- [63:03]
A
That's right. And of course now the housing market might be complicating that because the places with robust labor markets are very expensive. So it's a hard, hard, you know, the jobs have to be available and the people getting displaced have to have both the interest and the ability to get into those new trading programs. And that's a lot to ask. And historically the track record has been worse. But there, but there have been some like Like Trade Adjustment Assistance. That's this program where you have to be certified that you lost your job to import to get. And what you get is living money over and above unemployment insurance, plus extra training money. Now, for a long time it looked like that program was a failure, but eventually it adapted and, and the most. And of course the program didn't get reauthorized in 2022. A little bit unfortunate, but the last bit of research from Ben Hyman at the New York Fed said it was actually a much more successful program than it had been previously. Yeah.
- [64:06]
B
So the, the bad reputation is perhaps undeserved. Deserved.
- [64:10]
A
Well, in its current form, it might have been more deserved in some of its earlier forms. You know, he had a number, something like every worker who goes through the programs gets a $50,000 benefit over the next five to 10 years, which is a pretty good rate of return. And that that's even before you adjust for inflation. So, so there are programs, you know, there's another good study of, of unemployed workers in the state of Washington who went to community college. And Lou Jacobson and a few, and Lalonde and Sullivan did that paper and they found that it wasn't spectacular training, but it was pretty good. These workers got on average 10, 15% bumps. Now, there weren't a lot of older workers getting that retraining, but we do have some examples of places. And part of the problem is you have these spectacular training programs out there today like, like Per Scholis and Jewish Vocational Services and often in very specific industries. The problem is they're awesome for a small number of people who get to.
- [65:16]
B
Training, there's awesome micro programs. There's not awesome macro programs. Is that fair enough?
- [65:22]
A
That's right. I think A, those programs are expensive, B, they really demand people who are very ready to do technical work and technical training. That's not true of anybody. So you don't, as you said, you don't find spectacular programs at a more macro level. But you look at what the best community colleges are doing in the country and the ones that have figured out which, how to partner with local industry, they're pretty good. They're pretty good people. But again, if you're a 45 year old, it may not be for you and it may not solve your problem, but maybe if you're 30 or 35, a little more mobile, it might be better.
- [66:00]
B
Amazing. Well, Harry, we've covered a ton of ground. I want to thank you so much for, for coming on the podcast and sharing your expertise from this incredibly wide ranging conversation. You really held up to some, some wild speculation on my part and I am grateful for that. If, if folks want to follow more of your work, how can they find you?
- [66:20]
A
You can go to my website at Georgetown University. Just Google me and and you'll get links to my, to my Brookings papers, my Georgetown papers, et cetera. And I welcome people to do that. And thanks to you, Greg, for having me on. It was a really fun and stimulating conversation. Thank you.
- [66:37]
B
All right, well that does it for this episode of the AI Policy Podcast. We'll be back soon and look forward to it. Thanks. Thank you. Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mann. See you next.