The Evolving Impact of Robots on Jobs
29:51 Share

Hello and welcome to Skynet today's Let's Talk AI podcast where you can hear from AI researchers about what's actually going on with AI and what is just clickbait headlines. We release weekly AI news coverage and discussion as well as occasional interviews such as today. I'm Andrey Krennikov, a third year PhD student at the Stanford Vision and Learning Lab and the host of this episode.

In this special interview episode, we'll get to hear from the authors of a recent paper, The Evolving Impact of Robots on Jobs, Professors Chung and Yong Lee. Professor Yong Lee is the SK Center Fellow at the Freeman Spogli Institute for International Studies. He is also the Deputy Director of a Career Program at the Walter H. Shorenstein Asia-Pacific Research Center, a faculty

of the Stanford Cyber Initiative, a faculty affiliate at the Center for Global Poverty and Development and the Center for East Asian Studies at Stanford University. His research is in fields of labor economics, technology and entrepreneurship and urban economics, with his current research examining digital technology and labor and focusing on how new technologies will affect labor and how societies react to new technologies.

Professor Jong Chung is an assistant professor at the Economics Department of Auburn University and got his PhD from Stanford University in 2019. His research interest is in the field of international trade. Thank you so much, Professors Lee and Chung, for joining us for this interview episode. Thank you very much for having us.

Alrighty, well, let's go ahead and dive in. Our focus will be on your recent paper, The Evolving Impact of Robots on Jobs and Its Conclusions. I actually came to see this research by an article from the Stanford Human-Centered AI Institute that was titled, Why Robots and AI May Not Herald a Job Apocalypse, which is kind of letting us know what the conclusion of that paper might be.

So before we dive into any details, I'm going to ask both of you, can you kind of give us a high level summary of the project, of the research and why you found it interesting to work on it personally? Okay, sure. Maybe I'll go first on this one. So this project started actually quite a while ago. It's been about four to five years.

And back then, I was interested in comparing the effects of globalization on labor markets and also the impact of technological change on labor markets. Because those were two forces that scholars were debating and discussing and examining in terms of having long-term consequences on jobs.

And back then, John was a graduate student in the economics department and he was specializing in international trade. So it was perfect that I got in touch with him and he had the expertise in examining trade exposure, robotics exposure, etc. And that's how we got together, how it started.

And while we were examining this, we found that what interesting aspect was that the effect of robot exposure was changing over time. And we gradually shifted the focus of our paper to focus on this dimension that how robotics affects jobs, but its effects are changing over time.

So to just give you an overview of what the result is, there was already a well-known paper by MIT scholars that showed that robot exposure negatively affected jobs in the U.S.,

And we were finding similar results when we were examining similar periods. But as we looked into the data more, we found that that effect changed over time. That initially during the periods that they examined, early 2000s, yes, indeed, robot exposure reduces jobs, local employment and wages in the U.S.,

What we found was that in more recent periods, post-2010, more recent years, that robot exposure actually led to an increase in jobs in the local economy and also some uptick in wages as well. So that's sort of the high level overview of what we found.

I see. Yeah, that makes a lot of sense. John, maybe you can add a high level summary of how you went about the project and any personal angle on why you found it interesting yourself.

Yeah, so as Young mentioned, the project initially was going to look at the effect of the trade shocks on the job markets. And in many ways, the approach, the general approach we took in this paper resembles the way you might examine the effect of trade on the labor market. There is a very well-known paper

by MIT scholars, different MIT scholars. They look at the impact of the Chinese imports on the US labor market, which similarly found a negative impact so that the employment as well as wage went down because of the Chinese imports. And in many ways, the concerns about the job losses due to the technological innovations like robots and AI resemble the concerns about trade and how it affects the

employment. So I thought studying and thinking about how the robots affect employment and wages also helped me better understand how the labor market responds to different economic shocks, including international trade. I see. And so you say that it's similar to how you might conduct international trade research. Maybe that's a good starting point to talk about

What data did you use to actually analyze this impact? How can you tell the impact of robots on jobs in particular sectors or areas? So our main source of data for measuring the robot exposure is the International Federation of Robotics, which is an organization that, among other things, provides the number of industry-specific robots shipped to each country in a given year.

And I want to be clear here. So when you say robots here, we mean the industrial robots, which are required to be reprogrammable and multipurpose. We complement these robot data with the EU CLAMS data. This data set provides number of workers for each country in different industries and years.

So these two datasets allow us to measure the number of robots per workers in different industries for different countries and years. Along with the global exposure measure, we also use the census and the American Community Survey data. And with this dataset, we measure the employment and wages of different industries and regions within the U.S.

This allows us to calculate the industrial composition of different regions and also to measure the employment and wage growth of these regions, which are the objects of our interest. There are a few other data sets we use, like the trade data from UN Comtrade that we used to account for the Chinese imports and its effect on the employment.

And much of the effort was really to combine the different data sets so that their industries and regions agree with each other and remain consistent over time. And except for the robot data from the International Federation of Robotics, all our data are publicly available. I see. Very interesting. Certainly, as someone working in AI, getting a glimpse into how stuff is done in economics is pretty interesting.

I just wanted to add maybe one point is because, you know, a lot of scholars, economists have been examining this question of what's the impact of robots on jobs.

And the effects will likely differ across data sets because the data comes from different countries. Different countries have different institutions, labor unions, regulations, et cetera. So when we're thinking about examining whether an effect is evolving over time, we wanted to make sure there were a bench line that we could actually compare this to. And the main bench line was like –

The paper that we discuss, the scholars by, it's Asamoglu and Restrepo, which examined what's published and shows that robots reduce jobs.

in the US. So we wanted to maintain a consistent framework that we could actually compare that, yes, it is true, we find a similar effect using very similar framework and data, but it evolves over time so that the effect we're finding is not attributed to different institutions, different countries or contexts, but within the same framework that we're finding this evolution.

I see. Yeah, that's of course an interesting finding. I think it's a bit counterintuitive that at first robots seem to be bad for jobs and then maybe less bad and maybe even good. So can you give us a high level explanation of how you analyze the data, how you came to these conclusions? What I do is we exploit the different regions in the United States have different experience, different exposure to the growth of the industrial robots.

So some regions within the US are more heavily exposed to the impacts of robot growth because they are focused on the sectors that adopt a lot of robots like automobiles and electronics.

So what we find is that in the 2000s, these highly exposed regions experienced greater employment and wage loss or smaller employment and wage gain compared to the regions that are less exposed to robots. So that finding suggests that robot adoption causes smaller employment or loss in employment. And this finding again is consistent with the previous results that Young has mentioned.

But we used exactly the same design, but look at more recent years, and then there we find the pattern has changed. So the high exposure regions, the regions that are more heavily focused on the sectors that adopt more robots, actually saw a growth in employment relative to the other regions that are less exposed to the robot growth.

So our estimate suggests that between 2005 and 2011, each robot was associated with about 20 to 50 jobs lost. Whereas in more recent period, each job actually added about 15 jobs. I see. That makes a lot of sense. And then given that, maybe we can talk a bit about why might this be happening? What are the causes for jobs losing?

for robots' impacts on jobs to change over time? Yeah, so I guess the reason why we see this rebound effect that we believe is, so robots themselves change. And initially when robots were introduced, the quality of robots back in the 70s, 80s and 90s are different from robots today.

One can imagine even for the same robot that performs the same type of activity, for instance, welding. Welding robots probably today are much more efficient than they were back then. So what that does is that robots become productive in themselves.

And as robot, a type of capital becomes more productive, the firm becomes productive, which thereby can actually increase overall product labor productivity and labor demand by the firm. So this channel is what we believe is going on and so-called robot deepening, basically the same type of robots actually performing better or becoming more productive.

Another angle that we see anecdotally is there's these type of robots that are being created that are intentionally dangerous.

to augment humans. It's called collaborative robots, augmentative robots, different words. But basically, for instance, in construction, some tasks people can't do by themselves. They wear these robots and they perform the hard tasks, difficult lifting activities, whereas a human can focus on the maneuvering, etc.,

So we've seen the rise of collaborative robots in various industries, especially in automotive and electronics as well. And we think that's two channels where the robot augmentation effect might be happening.

I see. So that sounds like a lot like the commonly used bank teller story. So in AI, we often discuss the job losses and it's often pointed out that ATMs, our technology often results in some jobs being lost. But then because ATMs make banks sort of more productive, there's more bank branches and there's more overall jobs as a result.

Is that the idea of robot deepening, that robots become so useful that you can do more manufacturing and as a result there's more manufacturing jobs? Okay, I can try. So that is definitely a possibility.

So we do not quite see the details of the jobs that gets created or destroyed with our data set. So while we can suggest different mechanisms through which the robots are creating jobs, we cannot be exactly sure whether this job creation is due to new task being created, like these people managing the robots and taking care of the robots versus the other effects like the productivity improvement,

due to the robots? Yeah, I think there is some parallel to that, but there is also some differences. The case of the ATM is basically it freed up Tesla's time so that they could focus on different tasks. And that is true in some aspects, but sometimes, like in the case of robot deepening itself, it may be the case that...

It's doing the same type of activities and it's just performing better. So in the same amount of labor input with a better capital, maybe performing better. So that's one angle. The other is something that you just mentioned that, oh, by robots performing certain tasks, it's freeing up more time. In that case, what's the main difference is the range of tasks that the robots perform can be expanding.

And then that can free up labor's time and they can actually focus on different tasks that they have to compare to the advantage in.

Another angle that we haven't discussed yet is that in general, when these new technologies are developed, organizations need to change as well and adopt to actually become productive. And this could have been happening as well when there are certain types of robots being introduced or

organizations, firms might not be using them very efficiently because they don't know how to integrate it into their processes yet. But as they learn over their years and they find the right balance, they can become more productive as well. So there's this sort of time lag and productivity gains when new technologies are introduced. And this was the case

for IT, even electronics, electricity back then, as well as probably the newest technologies like AI. I see. Yeah, that makes a lot of sense. And this also makes me think of when talking about jobs and AI, often, you know, even just normal researchers in AI like me who aren't that diverse in economics,

There's kind of a story that with new technologies of AI, there'll be new types of jobs. So you'll need people who program the robots or otherwise, you know, there's going to be new roles for people to fill. So did you look or is that one possible reason for change over time or not so much? John, you want to take a stab? So I guess to add on to that, I think one of the reasons it's so hard to speculate these things is

the type of technology that's disruptive are precisely the kind that's difficult to predict. So I think our research actually indicates that as technology becomes more mature, it tends to iterate on the existing technology. So for example, robots, as the technology for robots are maturing, we see the new robots are replacing the existing robots rather than humans. And therefore that we see the

So we're benefiting from the productivity gain without this disruption on the employment. But of course, it's possible, especially with the technology like AI and other fields. These new innovations that's difficult to forecast now can be very disruptive and disruptive.

Yes, so going back, it is precisely the type of technology that's difficult to predict that will be most disruptive. Yeah, just to build up on that is, so I think, so what you're referring to is what's basically often referred to as a reinstatement effect that, oh, when cars were introduced, they're now taxi drivers. When spreadsheets were introduced, there's now these type of accountants that can focus using software.

And the idea is, yes, when there's these new AI technologies, we hope and maybe the case that there may be new types of jobs being created that did not exist before. So the creation of new jobs would definitely result in new technology resulting in more jobs.

But is this the case that we see here? We don't think that's the story here in our paper when we're looking, focusing on industrial robots and finding the effects primarily in manufacturing because economists are still examining this. They're trying to look into data to find like, are there any new types of jobs happening? And it's a bit slow. There's not good data on this.

But we're not in general, economists have not found new jobs being created in the sectors that robots are impacting the most, for instance, manufacturing. Most of the new jobs, if they were created, are in the service sectors. So personal health, healthcare.

finance, etc. So we believe that what we're finding here is more of the productivity related effects rather than the reinstatement effect. But as you mentioned, the case of AI or other new robotics technology down the road, definitely we do think that reinstatement effect can happen, but it's still at a very early stage that we don't have much empirical evidence to confirm new jobs being created yet.

That makes a lot of sense. And yeah, I think it is probably interesting and worth noting for AI researchers out there or people interested in AI. It's a little more complicated than the conversations we often have when we discuss these things. And so things like robot deepening and collaborative robotics are also interesting and worth being aware of besides just new jobs being created.

Now, if you can zoom out a bit and maybe get a bit more speculative, I think it's very interesting to see the results you have. And, you know, for the data that you have, it's factually what it is. Can you make an educated guess as to is this pattern likely to continue? Can you see looking forward into, let's say, the next decade,

Do you have any guess as to the impact of robots on jobs going forward? Yeah, it's always a bit challenging to make speculation. But so the results we're finding here is at least it's promising that in certain sectors where there are productivity gains from robotics, that sector expands and it hires more.

And even potentially there's a spillover effect where, oh, you know, the regions, one part of the region's economy, this manufacturing sector is doing very well. It could be a new manufacturing sector. Then, you know, the service sectors nearby will benefit because there's demand for service goods.

which could actually increase jobs in other sectors as well. So that's the optimistic view that we see. And we do find some of that evidence here. But to be cautious that, you know, if defective is evolving over time and robots are evolving, we can even imagine robots evolving even further, right?

where they become even more advanced, where we talk about collaborative robots. But suppose that robots become sufficient enough that they can perform a whole range of physical activity that humans could have performed.

In that case, it could be, you know, another dip where these super advanced robots take over a large chunk of the human's task and potentially result in, you know, reduction in jobs. So, again, that's very speculative. It's hard to know where technology will develop. But one thing that's interesting is, you know,

Firms and people have initiative in how to design robots.

So the way collaborative robots were somewhat designed was that they realized that rather trying to design a robot that, like for instance, certain companies, this was a case in Mitsubishi in Japan, they were initially trying to think of a robot that would actually replace construction workers, but learned that that wasn't very productive. And they intentionally designed to come up with robots that can actually augment humans.

And also when there are sectors where there's a shortage of labor, they want to augment these workers so that can actually facilitate them coming into the sector and maintain them in the workforce. So I think there's this aspect of, oh, it's robots are exogenous. No, it's not the case. Robots are actually endogenous and people can actually design the way that robots will function in the economy.

I completely agree with Yong, and I think more broadly, there's definitely room for the policies to help those who suffer from this type of innovations and other shocks. So when you do research like this, it's easy to forget the actual pain that's been felt by the workers and many workers.

And I think there's definitely room for policies to provide generally broader safety net to help the people, even if this loss is short term, to help them out and to insulate them from the types of uncertainty we discussed today. Yeah, so, of course, hard to speculate. And many people often kind of

armchair economists and predict very confidently that AI will cost all our jobs. And in my experience looking at the area, it seems like at the very least you can say that it's hard to predict and there's different kind of effects going on.

So maybe for our last question and to round things out, let's zoom out a bit further and just kind of talk about for AI more broadly, not just robotics, but also computer vision, language recognition, all these technologies. Should we kind of be mindful of similar effects that could be happening with technologies? So, for instance, you've mentioned, I think, that there's a delayed productivity gain as organizations kind of learn how to use these tools and integrate them

and make use of them and the technology itself develops for generally other AI technologies. Should that be also a factor people are aware of in thinking about the impact on jobs and the economy? Yeah, definitely. AI is even even it's it's the so-called the newest general purpose technology.

So robots, for instance, have been around, especially in the industrial sector, for a long period of time. In the case of Japan, which adopts robots very intensively, it's already declining in terms of adoption. So, yeah, AI is the hot new technology out there. And I think, yes, it's the same similar story that right now we're not sure, like, why

First of all, there's not very much data on AI adoption, how they're using this within businesses at the micro level. So despite economists being interested in this question, there's no really good data to examine this. And

There is speculation using different methods that AI could replace certain type of workers, certain type of tasks. But some evidence out there says it's much more widespread compared to robots. And it's across all spectrum from blue collar to white collar workers.

But I do think there is this delayed, potentially delayed productivity effect that will happen because I have this other paper that looks into AI adoption in the finance sector. And what we're finding is that looking into different occupations, AI does decrease productivity.

the jobs for what we find lower level, low skilled workers in banking and finance. But in general, it seems like AI adoption is highly correlated with the increase in managerial position and analytical position, not just tech positions, but overall.

So I think we're still it's a bit early to tell, but like similar to a lot of these technologies, initially, we might not find much productivity gains. But over the long run, I'm hopeful that these could result in productivity benefits and and actually increase demand for certain occupations.

But the caveat is that it could be like what we're seeing in the economy right now, talking about the K-curve recovery, that certain spectrum of the occupation, especially skills or analytical, might do better than those with less skill.

And the key aspect going forward is how are we going to train a set of workers without the set of skill that's suitable, that's suited for the AI economy? And that's not an easy task. It's trying to train people. The code itself is not easy. But that's something I think we need to keep in mind as we consider the impact of AI on jobs.

That makes a ton of sense. And that's one message you've tried to actually highlight that you found in economics that, you know, maybe we shouldn't be so dystopian and pessimistic about all jobs we've lost, but we can try to provide safety nets and think about the impact of AI on jobs. And so I'm glad you pointed it out. And certainly something I try to convey to people when they ask about this topic.

So with that, we'll go ahead and wrap up. Thank you so much, Professors Yong Li and Jong Chung for being on this episode. Thank you very much, Andrei. And thank you so much, listeners of this episode for being with us on the Let's Talk AI podcast. You can find articles on similar topics to today's discussion at scannettoday.com. Subscribe to us wherever you get your podcasts and don't forget to leave us a rating if you like the show.

Be sure to tune in to our future episodes.