Speaker 1: Hi, this is Bernard Marr. I'm here with Carl Frey at Oxford University and what I want to talk about is this very exciting topic of the future of employment. We are currently experiencing a new industrial revolution, new technologies are coming in to change our world of work and Carl, you together with your co-author Michael Osborne wrote a very influential paper which is now many years ago, was it 2013, is that right? That's right. Yes, so one of the most quoted studies in this field when it comes to the future of employment, maybe you can, it's very much quoted but also very misquoted. I think there's some scary numbers in there, they're high headline figures that is often talked about is that the study predicted that 47% of current, of US jobs could potentially be automated so maybe we can talk a little bit about that study. Maybe what motivated this, the key findings and what we can learn from that. Sure yeah, so what we did

Speaker 2: back in 2012 or when the study was published in 2013 but I think we started writing the pages straight because I was a bit more excited to look at the results I think it might have been more interesting and interesting in the future. And I'm very excited to be here with Carl and I'm paper in 2011 actually. So what we did is trying to look at how the current pie of employment, how exposed that is to automation as artificial intelligence, mobile robotics, all of these technologies become more pervasive. And the economics literature at the time had this dividing line that machines have a comparative advantage in things that are routine and rule based that can easily be specified in computer code from the top down. But at the time we were seeing a lot of examples of computers doing things that were commonly deemed non-routine like translation work, potentially driving a car, medical diagnostics, generative adversarial networks, dreaming up fashion models from thousands of pictures and that sort of stuff. So there was nothing that... Most people would deem routine. So we tried to figure out what this actually means for the future of employment and in terms of looking at the potential scope of automatability from a technological capabilities point of view. So obviously there's a lot of factors that shape the pace of automation and the extent of automation. So when Nissan produces cars in Japan, it relies heavily on robots. When it does the automation, it relies heavily on robots. When it does the automation, it does the same thing in India. It relies heavily on cheap labor. And even if Google Translate becomes perfect tomorrow, you still need a certified translator for certain documents to be valid. So unless you certify Google Translate, it's not going to replace the jobs of translators and so on.

Speaker 1: So the headline figure then though is 47, where there is the possibility of the existing jobs to be automated. But what it doesn't say... What it doesn't say is the jobs that were created, right?

Speaker 2: Exactly. So on the one hand, it only looks at the potential automatability of jobs from a technological capabilities point of view. And obviously other factors as well, like the relative cost of capital and legislation and consumer preferences, and even potentially worker resistance, shape the pace and extent to which automation will actually happen. And secondly, it looks at the pie of employment at one point. Yeah. At one given point in time. Yeah. It doesn't consider, you know, potentially emerging jobs. And if I were to go back and ask my great-grandmother if I could travel back in time and ask, what do you think that your great-grandchildren are going to do? She would probably not say that, you know, my son is going to be a hot yoga teacher and my daughter is going to be a software engineer. I think that's quite unlikely. In a similar fashion, we're in place today to actually try to predict what the jobs of the future will be. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. But I think one thing that I do still find concerning is that if we look at what machines are good at, they are especially good at doing many tasks performed by low-skilled, low-income workers. And we've already seen this hollowing out of the labor market, middle-income jobs disappearing. Okay. And it seems to me that looking forward, I think we're going to see a lot more jobs that are less looking forward at many of these sort of safe havens for low-skilled workers working as receptionists or security guards or telemarketers. Many of those jobs are likely to disappear going forward. And for people that don't have cognitive skills, I think that it looks less likely that those people are likely to prosper in the future labor

Speaker 1: market. Okay. So if you... One of the questions I get asked a lot is, you know, what is the future of the labor market? Yeah. And one of the questions I get asked a lot when I do presentations is, how do we prepare for this? How would you... What advice would you give to young kids today, to people in the labor market maybe that face some risk of their jobs being augmented and replaced by machines? What would you advise those people to do?

Speaker 2: Yeah. So in the end, I'm an academic. I'm not a career advisor. Okay. So I usually give them that answer. But clearly, I mean, if you look across departments, right, so in engineering science, we can't keep people that do machine learning for their full PhD because they, you know, get hired by Facebook and Google and get great jobs. So obviously, if you think that data is the new oil, which is clearly, you know, cliche, but it's a cliche for a reason, you would think that working with large data sets and doing machine learning and that sort of stuff is going to become a skill of value. And that's good for people that leave school now who can, you know, study information engineering. The harder part is for people that lose their jobs later in life. And I think we need to ask ourselves, doesn't it make sense for somebody at the age of 64 to sort of retrain from completing a new job? Or should we rather think about, you know, subsidizing their employment to some extent? Yeah. If they drop down from middle-income jobs to low-income jobs, government could potentially top up some of that income differential and create more incentives, better incentives to work and also reduce levels of inequality. And I think education is not going to be the answer to everything. I think it's part of it, but it's not the entire answer.

Speaker 1: So you said a key is the money. It's the middle-income people where you might don't have a college education. They are quite at risk. From your current experience, which kind of jobs would you put at the high-risk category for automation and which ones would you place at the lower level?

Speaker 2: So, I mean, so we found that quite a significant share of jobs are exposed to automation. And it's not like, if you look to the past, it was mainly back office work or production jobs in very structured environments that were exposed to automation. Now we find that everything from jobs in retail, transportation, logistics, construction, sales, the sort of potential scope extends across almost every industry and domain. Mm-hmm. Yeah. And if you think about, for example, in the United States, there are 3.5 million people still working as cashiers. If you go to an Amazon Go store, you won't see a single one. Now, that doesn't mean that those people won't find new work, but it's, I think, the sort of tendency, and we do find, see a very sort of strong negative correlation between skills, income, and jobs susceptibility to automation. Yeah. Yeah. It is, on balance, the low-skilled, low-income jobs that are most exposed to automation. And we also know from past studies that those are the people who struggle the most to adjust. Mm. So if you, in terms of the job skills that you find are least susceptible to automation, what are some of the key examples there? Right. Yeah. So the basis of our study is that we actually consider, you know, a lot of the jobs that are most susceptible to automation. Right. Yeah. We actually consider three key bottlenecks to automation. So we're saying that despite all these forces in artificial intelligence and mobile robotics, what are the domains in which computers still perform poorly? And one such domain is clearly complex social interactions, which were highlighted in the paper. And I think it was a year or two after, so we had to argue that no chatbot had performed well in low inner prize competitions. Those are some people who I know worked well in local enterprise competitions and, in essentially during test competitions where Shatterpots tried to convince human judges of them being a person. And I think two years after we published our study, there was a breakthrough and one Shatterpot actually managed to convince 30% of human judges of it being a person. But it did so by pretending to be a 13-year-old Russian orphan boy speaking English as his second language with no understanding of English culture. And this is basic text communication, right? So even if computers become perfect for text communication, there are still so many in-person type of interactions that are hard to automate where we try to motivate our colleagues and persuade somebody that we're right about something and so on. So that's one bottleneck. Another bottleneck has to do with creative type of work. There's a big debate in machine learning. A community as to whether computers can be creative. I think that most people that say that they can or are, or should never say that they can't be, if you don't think that we're anything else than just a composition of atoms, you should think that it should potentially be possible at some point. But I think it's quite far off. And I think the people that argue that computers are creative already tend to conflate novelty with creativity. So I could draw something on the wall here and call it creativity. I could draw myself an artist, right? But you would be quite unlikely to buy my painting. And the hard part is to arrive at something that is novel, that actually also makes sense. And novelty is not the tricky part. And the last bottleneck has to do with the perception and manipulation of irregular objects and navigating complex environments. So for us, it's quite easy to distinguish between an important document on the floor and a piece of rubbish. But so conceptually, if you think about having an automated cleaner, it's actually not that straightforward. And other things like a pot that is dirty and needs to be cleaned, and a pot that holds a plant to toss, actually not so easy to explain. And I think one thing that we argue in the paper is that the automated cleaner is one of the last things we can...

Speaker 1: I guess so. So cleaners, plumbers, I think are a great example where you need the dexterity, the understanding of complex environments, where you have a new job challenge every day. Yeah, very good. Very good. Thank you. That was useful. Thank you very much for your time. My pleasure. Thank you.