Artifical Intelligence & the Future of Work, with Carl Benedikt Frey (Oxford University)

Hello everyone and welcome to our The Week That Was in Europe podcast. My name is Dirk Schumacher. I am the head of European macro research at Natixis, and with me as usual is Claus Adam, professor of economics at the University of Mannheim.

Yeah, hello everyone. So today we'd like to discuss about the future of work. With our special guest, Professor Carl Benedikt Frey from the University of Oxford. Carl Benedikt is the Dieter Schwarz Associate Professor of AI and Work at the Oxford Internet Institute and also the Director of the Future of Work Program at the Oxford Martin School. So he's an economist by training, but also a historian and has written extensively on the role of technological disruptions for labor market outcomes. And his recent book titled The Technology Trap has received numerous awards, including the Financial Times Book of the Year Award. So we are very glad to have you with us today, Carl. So welcome to the podcast.

Good to be with you.

Yeah, I think we have all noticed that the onset in particular of AI and AI tools is gonna change how things are gonna be done. It's another big technological revolution. And already artificial intelligence has made significant inroads into areas that were previously almost exclusively reserved for humans. And this includes areas such as now conversational interaction, prominently with ChatGPT, but also prediction and recognition patterns and tasks, also writing and teaching tasks for that matter. So they will be competitors for us at universities. And the questions that everyone asking now is of course, what implications will these new AI technologies have for labor market outcomes going forward? So you, in your recent book, you suggest that we can learn about the likely future consequences of AI technologies by looking at past technological revolutions. And you argue that the near-term implications of these technological innovations are almost always negative in terms of their short-term labor market outcomes, perhaps not so short-term, perhaps even medium-term. So, and, uh, this is true even though, of course, in the long term we are gonna get long benefits, big benefits in the long term. So how do you see the situation now with these AI tools and the AI revolution upcoming? Will we have again negative labor market effects from these technologies in terms of, say, hollowing out middle-class jobs and so on?

Yeah. So what I argue in my book is that there are essentially 2 types of technologies and drawing upon a model by Daron Acemoglu and Pascal, uh, Trepo. I think it's sort of a helpful distinction. So for example, the automatic elevator, all it did was eliminating the operator, uh, there was nothing much more to it. Uh, and if you think of a technology as the telescope, all it did was allowing us to look at the moons of Jupiter and, uh, explore, you know, what's out there in the universe. It didn't really replace, um, any workers. So there's some technologies that replace us in existing activities, there are some uh, technologies that augment us in existing activities and, uh, even create new tasks for labor. And then there's a lot of intermediate cases. So if you think of, uh, technology like the internal combustion engine, well, it replaced the horse carriage, but it also created a sort of vast industry around automobiles and road commerce and various supplier industries, and you name it. And, uh, empirically And I feel fairly confident in saying that the automobile was more of an enabling technology than it was a replacing technology. And I think a key question surrounding AI, is it replacing or enabling technology? And I think, like the internal combustion engine, it's a little bit of both. It's already, you know, creating new occupations. So prompt engineer, I hadn't heard of a year ago.

No.

It's all over the place. At the same time, though, the jury is still out. But I do think that as it's been used so far, at least, it's mainly being used for automation. And many of the applications that we've been sort of looking at over the past decades, some of which I should say have not materialized yet, but I think are likely to materialize, like autonomous vehicles, are decisively sort of automation technologies. And, and I think the same was true to a large degree of the first industrial revolu— uh, industrial revolution, which didn't create so many new industries. Um, um, and it did create some new activities for labor, but not that many. It mainly sort of related to the automation of existing industries like textile production, which drove down the wages of textile workers, replaced a host of jobs in, uh, domestic industries, uh, put downward pressure on people's, um, incomes. And I think that's something we've seen with the computer revolution playing out since the 1980s in frontier economies like the United States in particular. Um, and I think it's something that, uh, is likely to continue, uh, with, um, artificial intelligence. So one of the key implications of our 2013 paper, if you buy those estimates, is that it's primarily low-skill, low-income jobs that are most close to automation. I think there's, you know, got to be some debate around that, especially in the light of generative AI, uh, more recently. But in my view, uh, those estimates, uh, still ring true, right?

So you'd say people that work, say, in call centers are facing now a risk of being replaced by language models?

Or people working call centers, working as telemarketers, a lot of the occupations we looked at in our 2013 paper that are automatable related transportation logistics. I think if you look at, you know, mining, we see a use of autonomous vehicles there, and we see autonomous cargo trucks being used at harbours. We haven't seen the adoption of autonomous vehicles to the semi-green, unstructured environments and public spaces. I think it's probably a matter of time, but there will be some disagreements around that. And we're also seeing now the use of, you know, generative AI for, you know, things like replacing fashion models. One of the sort of more surprising estimates of the 2013 paper was that fashion models were exposed to automation according to those estimates. And a lot of people ridiculed that. And, you know, there are probably things you can ridicule on both sides. But, you know, I think that, that turned out to be correct.

Yeah, film industry and all sorts of other entertainment industries also, actors are at risk now because they can be replaced. That's quite surprising perhaps, yeah.

And disheartening. I just was trying to move into this acting world, but now, okay, I'm stuck with economics. Anyway, you already alluded to some of the sectoral effects, but let me ask again a bit more specifically about the different sectors. So the research by Nir Jarmowicz from the University of California at San Diego, who argued that past technology revolutions mainly caused routine manual work to disappear. And one can think of factory workers or waiters getting replaced by robots or sectorial support functions in offices getting replaced by computer and software. If that's a template, you'd say that the new feature of AI is that it will also allowed to substitute routine, routine mental work. So routine mental work includes many tasks currently performed by accountants, tax advisors, and again, maybe university teachers. Would you agree with this viewpoint that would see determinants of working as being more relevant for assessing AI replacement risk?

Well, I certainly agree that in the past it's primarily been routine rule-based activities that lend themselves to automation. And when we wrote our paper back in 2013, uh, most models suggested that computers hold a comparative advantage in, you know, domains which can easily be specified in computer code and, you know, can have a programmer specifying in a top-down manner what the technology should do at every given contingency. But I think what's happened with machine learning is that, um, it allows us to automate things beyond Routine work, right? So there's no computer programmer in the world that can specify in a top-down manner what, you know, a human should do or what a car should do at any given contingency in city traffic, because there are just so many rare events that you can encounter that it's completely, you know, impossible to foresee it. So what we're trying to do instead is allowing the machines to infer the rules of the games. themselves by tapping into the data trails that we leave behind and trying to predict what the human would have done in a given situation. And that turns out to be also quite hard because there's a lot of tail events, there's a lot of things that can happen and that will not necessarily be in the training datasets. But at least it sort of expands the premises of what computers can do. So I do think that sort of routine rule-based activities, they lend themselves more to automation than other activities. But I think what's changed over the past decade is that, you know, we're now able to automate things beyond that. And then you can— it's a matter of definition, right? Is it that those tasks are becoming routinized, uh, you could claim, or is it that the sort of potential scope of automation has expanded beyond routine tasks? I think that's, that's basically, uh, just a question, um, of, of definition. And then there's sort of another aspect to it which I think has sort of always been true but is often forgotten, is that The way we automated a lot of activities historically was not by, you know, having a robot replicating exactly what a human would have done, uh, in that realm, right? So we didn't automate away the jobs of medieval craftsmen by building robots that did exactly the same motions, right? We didn't automate away the jobs of laundresses that, you know, uh, by building robots that walk down to a well and then did perform the sort of motions of hand washing and then walked up to the house again and, you know, hung the clothes somewhere to dry, right? We did that through the electric washing machine. So what many economists, I think, have done is trying to look sort of at task composition of existing activities and whether they lend themselves to automation or not. And I think that misses a lot of the action because often you can restructure those tasks, and many of those tasks will not be relevant. So, you know, A taxi driver needs to hold the steering wheel, but we don't need an algorithm to hold the steering wheel to drive a car, right? Um, so I think that's also one way in which often automation goes beyond routine activities, because we're able to restructure the environment in ways that, you know, make existing activities increasingly automatable.

So things that didn't look like being possible to automatize. You mentioned that now these new technologies learn the rules themselves, but in some sense success has been enormous, but also, you know, the tail events seem to loom large, especially if it's sort of sensitive activities like driving a car where you can hurt persons or can produce a bad accident, at least, you know, hurt the car. So are you very optimistic in terms of, you know, the AIs learning the remaining 1 or 2% of tail events that they apparently have still trouble with learning? at the moment? I mean, we have been promised this technology for 10 years now by Tesla. It hasn't fully materialized. And, uh, are you expecting it to come, or is there still some fundamental unsurmountable issue, uh, at the bottom of it?

I mean, I, I, I do, first of all, believe that in high-stakes environments, we're just less likely to adopt these technologies. And there's been a lot of progress in autonomous vehicles, just as there's been progress in large language models. But with large language models, it's not the end of the world if you produce some incorrect text. You can edit it, you can fact-check it. Um, but if it's sort of, uh, autonomous car hits the pavement and there's an accident, well, unfortunately, it's irreversible, right? Um, at the same time, it also seems that, you know, we are a lot more risk-averse than we were 100 years ago, right? When we let out the first cars in the streets and the first tractors were adopted, you know, some of them exploded. Uh, they had all sorts of issues. Uh, the first, you know, train ride, ever to take place in Britain actually killed a member of parliament. We still have trains today. So I think it's sort of also a question around risk appetite. And, you know, the training in cities is important for this actually to, you know, get much closer to 99.99999-something percent. I think we'll never get to 100. And it's important to remember that humans are not 100% either. But I do think that there will be incremental improvements that, that, that will get us there. But I, I think there are, as you mentioned, uncertainties around that.

Yeah. Now let's turn, uh, you know, our— let's look back now a little bit at, um, the computerization revolution that you've studied. So you mentioned back in 2012 you've written, uh, basically a paper, but also another book together with Michael Osborne that examined, uh, basically how susceptible different jobs are to being replaced. by computers that arrived increasingly in the workspace. And your analysis back then looked at more than 700 different occupations. You mentioned models as one of them early on, and you sought to determine the number of jobs that are at risk overall, and then how in the cross-section of people that risk is spread across the wage distribution or the educational attainment distribution. And then you estimated that about 47% of total US employment was at risk due to computerization, and that high educational attainment, which often of course goes along with high wage and therefore high wage, were factors that mitigated computerization risk at the individual worker level. So now with basically 10 more years of data since you've written that book, and looking at what actually happened, you know, how well did the predictions hold up? probably areas you were spot on. You mentioned modeling and other things, but perhaps there were also things, areas where, you know, the world evolved differently. So how would you say, did your prediction work out?

First of all, just let me push back a little bit on sort of the word predictions. In my mind, you know, our study is less of a prediction than, you know, taking the pie of employment at a given point in time and then trying to estimate how exposed the workforce, uh, at one given point in time is to automation, right? So we're not saying any anything about, you know, whether AI will cause more, uh, destroy more jobs than it creates, or any of the equilibrium effects— to use a bit of economic jargon— in, uh, that paper, nor are we talking about the pace of adoption of new technology. So we're sort of really looking at what can be done from a technological capabilities point of view. So to give you One example, in that paper, one of the estimates suggests that waiters and waitresses are exposed to automation, and we now have restaurants that are essentially giant vending machines. So yes, they are potentially automatable from a technological capabilities point of view, but that doesn't mean that consumers will prefer those type of restaurants. Ones. Uh, in similar fashion, if Google Translate becomes, you know, perfect— and it is very good today— well, if you want to, you know, have your university degree translated, you need it to be certified by a human translator. And unless we certify Google Translate, it's, uh, not gonna, uh, replace a whole, uh, lot of jobs. So I'm not sure that sort of evaluating our predictions based on, you know, looking at employment figures, um, is the best approach, but it is an important part of it. And, uh, one thing that I mentioned earlier that hasn't materialized is a lot of what happened in transportation, uh, in particular, right? So we don't have autonomous buses, taxis, uh, trucks on the road, and therefore in those particular domains, we haven't seen, uh, any dent on employment. And rather, through sort of, you know, uh, ridesharing apps and so on, there's probably been an increase, uh, in the number of people, uh, driving taxis, for example. Um, so I think, you know, you can look at these estimates, and if you think of them as, you know, are these activities automatable? in my mind, sort of the key bottlenecks that we outlined in the paper are, you know, still valid. And, um, I think for most occupations they sort of yield intuitive results, uh, sort of based, based on those engineering bottlenecks, which we can come back to. But if you look at sort of the adoption of those technologies and how they've shaped employment, I don't think that, you know, you used the sort of term those predictions, so I don't think it probably yielded a lot of predictive power in terms of replacement so far, but I do think, you know, when autonomous vehicles and things like that arrive, those predictions will, or those estimates will, uh, probably be more seen also in employment statistics.

So, but that past experience then probably also shows that assessing the potential for AI to replace is even— is more difficult given that it has the broader potential. And it's not as clear maybe as it was back then with computerization which jobs would be those most at risk, given that you just alluded to the fact that it's not so clear what the routine task is anymore given the potential of AI. Is that right, that if it has been already quite difficult in the past with the computerization to assess what implication will be. It will be order of magnitude, maybe even more difficult for AI.

I think it's really hard to assess, and it's no precise science by any means. Then, you know, uh, you know, some might argue that it's so hard that we shouldn't, you know, try to do it. But I think it does tell you something about which type of jobs that are most likely to be affected, what type of skills And those people have. I think it's important to try to get some ballpark in terms of the, um, number of jobs that are exposed. Is this 5%? Is it 35%? Is it 75%? I think that does matter. Um, in economics, like in any science, there's no paper that will shed light on everything. And, you know, there's literature that adds up to something, and often there are several different literatures, and add up to something. And I think the studies, um, you know, have sparked a debate that didn't exist to the same extent when we were writing back in 2013. So, you know, I do think this question is notoriously hard to answer, and that's one of the reasons we boiled it down to something like, you know, of the current fraction of employment, how exposed are jobs to automation? We didn't even try to sort of estimate how long sort of it would take for these technologies to be adopted, which is then shaped by numerous other factors.

Where there can be no disagreement, I guess, is that the potential for negative labor market effects is quite big. And that begs the question to what extent governments should, if at all, play any role in maybe shaping the direction or buffeting the adverse effects here. Are there any examples where in the past governments have done that successfully? If they tried to, or what's your take on what governments should do faced with this potentially quite negative outcome?

I think there are good theoretical arguments for governments to try to steer the direction of technological change. So if you think about it, if we go back to 1900, electric vehicles and gasoline-powered cars are essentially on par. And then you have huge oil discoveries, among other things, that sort of tilts the balance towards internal combustion engine. And, you know, a century later, we're trying to get back into electric vehicles. And you could argue that that was a harmful lock-in. And, you know, if we had had the foresight at the time to tax gasoline more, maybe we could have, you know, uh, uh, spared, uh, the past century of, uh, you know, very high carbon emissions, um, and, uh, got us into— self into better place. And you could argue something similar around, you know, uh, digital platforms and the advertising model, which a lot of people argue are harmful. So maybe if you, you know, use a tax on digital advertising, you can, you know, get firms to adopt a more subscription-based model. So you can, you know, like Darren Acemoglu and Pascale Restrepo have argued, you know, uh, try to adjust the tax code Um, so that we don't incentivize automation, right? If payroll taxes are very high and taxes on capital are very low, you're incentivizing businesses to invest more, um, in automation technology. So I think there's a good case for trying to steer the direction of technological change. I, I do think it's very hard in practice though, because we don't always know what the outcomes are likely to be. We didn't know a whole lot about climate science back, you know, Uh, in the 1920s. Um, and, you know, many of these AI technologies have very different uses, so we're not going to be able to foresee every single use case. So I do think it's hard in practice. I do think, uh, do think that incentives matter, and, uh, you know, the tax code might be a place to start, but I'm somewhat skeptical about our ability to actually right policy that, you know, steers the direction of technological change.

Yeah, perhaps a complicating factor is that the West is now in a sort of a position of strategic competition with China in particular, and who's also strongly pushing on these AI technologies and perhaps is even ahead in some of the areas. Now, it seems that with this situation, the appetite for regulation is perhaps not very strong, perhaps on the American side. Certainly not. And so is the West going to be forced to ignore potential side effects, um, in the AI race just to be staying ahead of the competition? Or, and what would you see are the potential risks with such a strategy?

I don't know, honest answer. So I mean, if you go back to the Cold War, I think we were reasonably successful in containing the proliferation of nuclear weapons, although there were certainly events along the way that were very critical. If you look at, you know, the proliferation of biological and chemical weapons, we, you know, seem to be able to contain that reasonably well for the most part, even during the height of the Cold War. So I think that lens gives me some optimism. But I do think you're right when you say that, you know, geopolitical competition makes it harder to achieve containment. And indeed, one of the reasons that Europe ended up being so dynamic during the sort of 18th and 19th centuries, and China was not, is that there was a lot of competition within Europe, and that wasn't the case in China. And of course, that did uh, you know, create a lot of innovation and progress in the West, but it also ended in 2 world wars and, uh, disaster. So I don't think that anybody knows, I can know the answer, uh, to that question. There are probably some domains where there is, uh, some common ground when it comes to, you know, uh, disinformation. Uh, for example, I don't think, you know, any country will want their population to, you know, uh, get bad medical, uh, advice from, you know, ChatGPT or any other technology. Um, so, you know, I think it will probably be something where we can go application by, um, application, but I think it's, it's, uh, Important question to ask.

Okay, we've been quite negative so far talking about the replacement, all that, but of course there are most very likely going to be quite a bit of positive effects as well. In particular, productivity growth is likely to be, at least in the long run, higher. Question is how much, and what's your best guess there, what we could expect from all this? Also, maybe in comparison to, say, the computer revolution, is that an order of magnitude more we could expect going forward on the productivity front?

Well, if you think that 47% of jobs are automatable, you would think that you should be optimistic about the future of productivity. At the same time, though, most productivity growth historically has not just come from automating existing activities, right? So if we just automated everything we did in 1800, we wouldn't be much more prosperous today. And a lot of the growth that we've seen is, you know, our ability to come up with new products, new industries, new activities, um, for labor that were previously inconceivable. And I think a question is, you know, will, uh, AI allow us to do that? It's arguably a method of invention, and it could arguably, um, accelerate science and has done so already in domains like protein folding. But at the same time, you could, you know, make the case, same case about computers and the internet, right? I mean, in many ways, those were sort of the ultimate technologies for research and innovation, and productivity growth has been fairly disappointing since. So I would expect to see an uptake in, you know, generative AI. I would expect AI to contribute towards more— towards greater research productivity. But I'm very unsure if it's going to be a very significant uptake. And I do actually worry a bit that we sort of got to use AI to do sort of more things that we're already doing, but slightly more efficiently rather than new things, right? So if you think about what generative AI does, for example, is rehashing existing things and finding new combinations in existing patterns. You can, you know, prompt it to write a letter in the style of William Shakespeare, But Shakespeare already existed. And if you want to generate something entirely new, what do you optimize for? And AI has been very good at doing things where the reward function is reasonably clear, like playing video games. Maximize game score is fairly straightforward. But if you want to generate something entirely new, what do you maximize for? And so I do— I'm cautiously optimistic about the future productivity growth in that light. But, you know, I'm far from certain.

It's certainly going to be the case that many professions are going to be complementary to these new tools, and hopefully that their productivity— or they can do things faster or better than they used to be doing. And now, so what can we now recommend with the advent of these new tools to young people? You know, what are the jobs that are going to be required in the future? What should they What should young people study? What sort of tasks should they be able to perform in a world where these new technologies are around? Any suggestions to them?

I think, first of all, I think we shouldn't just change everything we do in education. So ChatGPT is decent at writing. It's better at writing than many people, but that doesn't mean that we shouldn't teach people to write. So I presume it's the same for you, but what I've I just want to learn something and sort of collect my thoughts. The process of writing is actually really important to do that. So I think if we outsource the process of writing to ChatGPT, we're gonna lose a lot in terms of learning. So we should use it as a helpful tool, but we shouldn't, you know, we should still, you know, teach people to read and write and that sort of thing. And secondly, I think it's likely that AI will put a premium on in-person communication. So if, I mean, one way of putting it is that if AI writes your love letters and everybody else's, the first date becomes more important. And I think that's true in a lot of other activities as well. And we do know that there are some people that have the ability to walk into a room and they make the room. And I think that skill set is going to be even more important in the future. So that relates to the sort of second bottleneck we outlined in our 2013 paper. Complex social interactions are one of the activities that I think are still very hard to automate, especially in-person type of communication, of course. And I think that remains true today. And I do think creativity. So as we discussed earlier, right, sort of AI does quite well where you have a clear reward function, but there are many instances where you don't have that. And then finally, there's a lot of activities that are high stakes, navigating complex environments. I think the automated cleaner is still one of the things that we are likely to see in the near future, at least in homes that are not purpose-built for, you know, automated cleaners. So Those 3 sort of broad domains I still think remain relevant today when it comes to sort of understanding what type of activities that are likely to remain in the human domain and the skills that are relevant.

Okay, I'm going to try now a short summary of all that what you said, which as usual is quite daunting. So starting with the first question, will AI be negative for the labor market? You said very much depends on whether it will replace replace or augment existing labor or workers. And AI is a bit of both. But so far you would say it's mostly automation, though that doesn't have to remain the case. Is the replacement risk highest for routine mental work? Well, you said that AI is different than, say, previous waves in a sense that it will also allow potentially to replace non-routine work. And it's not clear anymore actually what is routine work when you talk about AI. And of course, The risk appetite, you said, the society as a whole has, will play an important role in deciding which parts can be or which jobs will be gone or not. And automatic driving is one example here where risk appetite seems low, at least compared to previous times. The past effects of computerization, you push back against those who are saying maybe what is your forecast there? You said it was not really a prediction you did in the paper, but more to assess ex ante what are the potential for replacements because of computerization to get an idea about the ballpark numbers and what is possible, but that does not mean it will go down that route. So very difficult to make concrete forecasts. The role of governments, you said at least theoretically there are good arguments why governments should try to shape a bit the way technology evolves, but highlighted as well, of course, Maybe that's in practice very difficult to make. With hindsight, you can say that would have been a good route taken, or it would have been. But again, when you're in the thick of the things, it's very difficult to say. Competition from China on all that, you said it's very difficult to know how that plays out, but alluded to the say cold war, that there's at least a potential for finding an agreement. But nobody knows the answer. To quote what you said, productivity and AI, it's not just automation that's needed, it's new ideas actually, which are better research, which, which makes us more productive in the end and imply prosperity. And there you're cautiously optimistic. And finally, what are the new jobs or what should people have in terms of skills? complex social interactions, in-person communication, and creativity, though obviously creativity is just very difficult to teach. I guess that will remain a challenge. Was that a fair summary of what you said?

Right on target.

Thank you, and thank you very much for your time. That was very, very interesting and a topic I'm sure will keep us occupied one way or the other going forward.