The following is a transcript of a Social Europe podcast in which Social Europe Editor-in-Chief Henning Meyer discusses the impact of the Digital Revolution on the nature of work and inequality with Michael A. Osborne, Associate Professor in Machine Learning and Co-Director of the Oxford Martin Programme on Technology and Employment at the University of Oxford.
Okay Michael, thank you very much for joining us today to have a discussion about innovation, technological changes and what it means for the future of work. Together with your colleague Carl Benedikt Frey, you’ve published several studies that have made a big impact. First of all, can I ask you, when you write about the changing nature of innovation, what in particular do you mean? What is changing?
Okay, so the developments about which I’m excited have emerged from my own field, which is that of machine learning, which many of your listeners might know better by the name of artificial intelligence. Our idea is that developments in the field of machine learning, accompanied by those in mobile robotics, are really going to have quite a transformational impact upon human employment. In particular, we point to the introduction of algorithms that are able to perform sophisticated subtle decision-making in a way that was previously the reserve of humans, that are able to navigate the streets, that are able to make recommendations of products to customers.
All of these things are very likely to substitute for human workers in the fairly near future. So, our study set out to put some numbers to this phenomenon. In particular, we used a data set from the US, in which we had 702 different occupational classes and for each of those we had a list of characteristics of the job.
So, this organisation in the US had gone about determining their requirements for different types of skills within each of its different occupations. These included things like persuasion, manual dexterity, originality – all characteristics that we thought might be predictive of a job’s susceptibility to automation.
So, on the basis of that data, we used, in fact, a machine-learning algorithm and the characteristics of jobs about which we were most confident were not going to be automated in the next 20 years, or that, in fact, we had already seen some evidence of being automated, to try and identify the characteristics of jobs that rendered them susceptible to automation and thereby to predict for all jobs a probability of automation. In particular, we arrived at this headline figure that 47% of current US employment might be at high risk of automation over the next 20 years.
Okay and if you want to categorise, in the international literature, there seems to be a tendency that the jobs that are less likely or the least likely to be substituted are ones based on creativity, even though there was a discussion about it, but also the ones, in particular, based on social capital and human interaction. Is this something you would also agree with?
We agree entirely with that characterisation. In our study we highlighted three bottlenecks to computerisation, as we call them. They were exactly creativity, social intelligence and perception or manipulation. So, what unites those three bottlenecks is the deep reservoirs of tacit knowledge that humans have about the task they’re performing.
To give an example, consider the generation of a hit song, a quintessentially creative task. It’s very easy to get an algorithm to churn out an endless stream of songs, but it’s very difficult to teach the algorithm the difference between a hit song and not such a good song.
Because, in doing so, as a human we draw upon this enormously sophisticated conception of our social environment and our cultural context, in a way that’s very difficult to explicitly state in code. So, it’s exactly those three characteristics of jobs that we think are most difficult to automate.
The German Labour Minister has recently commissioned a study that took a slightly different view of the automation potential. They have basically three main points of criticism of your methodology. The first one would be what you seem to be describing is more an impact of technology on ‘activities’, rather than ‘jobs’ themselves.
The same jobs might not even perform the same activity, so there is a difference between the activity level and the actual job level. The second one is that you estimate the potential that you rely on that’s too much based on expert opinion, which tends to overestimate the potential for substitution and rationalisation.
The third aspect was that even though there might be this kind of technological potential, there are also social, legal as well as ethical barriers to implementation, which would mean that the full potential will not be achieved. So, the figures they came up with is 9% in the US and 12% of German jobs might be at risk of automation. What is your response to that kind of approach?
Well, my response to the second of those criticisms first, because it’s the only one with which I really disagree. They highlight our use of expert opinion as a basis for our study, which is true, but I think more needs to be said about that.
The first point is that we weren’t slavishly following the advice of the experts upon which we drew. The way our algorithm worked, which is itself a quite sophisticated piece of machine learning, is that the advice of the experts, the labels they provided to 70 different occupations, as being either automatable or not over the next 20 years, were treated as noisy or untrustworthy estimates of what will actually happen in reality.
So, our algorithm is actually quite tolerant to mis-labelled examples. To give one kind of example: we included waiters and waitresses as an example of an occupation that was non-automatable and I guess the thinking there was that the kind of small talk that a waiter or waitress makes in the process of performing their job is not something that’s readily automated. That it’s very difficult to teach an algorithm all the social nuances that a waiter or waitress might draw upon.
Nonetheless, our algorithm had come back and told us that the probability of computerisation for waiters and waitresses was 94%, which struck us as quite high. But interestingly, in the time since we’ve published the studies, so since 2013, we have actually seen restaurants, particularly in the US, automating some of the jobs that waiters and waitresses might have traditionally been asked to perform, in placing tablets on tables in restaurants.
So, these tablets were able to take customers’ orders, they were able to recommend products to them, in fact, they were able to substantially increase dessert sales in particular, by as much as 30%, by providing customers with these enticing looking images on the tablet.
So, the point I was trying to make is that the algorithm we use wasn’t completely subject to the biases that the experts we used provided us. It was able to correct, as I say, for some mistaken labels. In particular, we performed some sensitivity analysis in which we sub-divided the training set that we used, so, in particular, we used a sub-set of 35 of the 70 occupations that we had labelled to try and predict the other 35.
We did this over many iterations of splitting that set into two. Yet, regardless of the split that we made, we found that we were able to very accurately predict that held out half, given the half that we’d retained for the algorithm. So, what this tells us is that, yes, it’s possible that we’re wrong entirely about what is automatable, but at least we’re consistently wrong, that there doesn’t seem to be a whole lot of inconsistency in this set that was provided.
Okay and would you say it’s basically that for some of the jobs you were actually on the cautious side?
I think that’s true, yes. So, for waiters and waitresses as one characteristic example, we’d over-estimated the influence of social intelligence, but what the algorithm picked up is that the low creativity score, or originality score for waiters and waitresses does indeed render them quite susceptible to automation.
Okay. What do you make of the argument that the likely impact is not necessarily substitution only, but it’s basically a changing face of the job, so that the job description itself will change and the human activity will basically be more augmenting technological capacity, rather than being in direct competition. Is this also something that you would subscribe to?
I’m very sympathetic to that point of view; I think that is indeed how automation takes root in the economy. So, historically, of course, we saw people like the typing pool of the 1950s replaced by word processing software, despite the fact that it’s not that the word processing software was able to do everything that the typing pool was able to do.
The job was simplified to the point that it was able to be automated. So, the job of what that typing pool contained was essentially eliminated despite all its characteristics not being automatable. But, more broadly, there was a study recently from Australia, which tried to put some numbers to this phenomenon by which the characteristics of a job might actually change over time and, in particular, how they might change in relation to automatability.
This study will be coming out of the office of the chief economist in Australia, it had found that there have indeed been quite substantial changes within occupations, as pertaining to their susceptibility to automation, even over the last ten years. So, I think that is a very substantial kind of effect.
I’m not sure whether the real world consequences will mean that occupations become either more or less automatable, I guess my first point was that, in fact, it’s possible to simplify tasks, simplify jobs to the point that they are automatable, just as jobs are able to be rendered more complex to the point that they’re not automated and which of these will actually play out in the real world is still up for debate.
So we’re basically talking about, when it comes to the impact of technology on jobs, three levels. The first level would be what kinds of jobs could be replaced by technology, the second one would be what kinds of jobs will change, maybe even fundamentally, because of the impact of technology and the third one would be job creation, what kinds of jobs will be newly created as a result of technology. What would you say to the third point?
So, again, a very important point we quite explicitly acknowledged in our original study is that it did not consider the prospect for new job growth. However, in subsequent work, my co-author Carl Benedikt Frey has indeed looked at what new jobs have emerged over the last decade within the US and finds, quite alarmingly, that the new industries that have been created now only contain about half of one percent of current US employment.
So, these new industries that have emerged haven’t necessarily generated a whole host of new employment and perhaps the most emblematic example is found in the tech sector, where a company like WhatsApp, for example, was bought out for 19 billion US, at a point at which it had only 55 employees, which compares quite unfavourably against a business of, kind of, similar valuation, like the fashion retailer GAP, which has about 137,000 employees.
So, yes, there are new jobs being created, but it’s not clear that there are as many of those new jobs as we might like and it’s also not clear that the jobs that are being created might be suitable for the people who are put out of work due to trends in technology.
So, if we look at the list of fastest growing occupations since the year 2000, they include jobs like data scientists, IOS developers, Android developers. I don’t really think that there are going to be all that many people put out of work by, for example, self-driving vehicles, people like truck drivers, who are able to move very naturally into these quite high-skilled occupations.
So we might also see, basically, a polarisation of the labour market? I mean, when I discussed my own work on the impact of technology on work on BBC Newsnight at the end of last year, Evan Davis, the presenter said, “Well, we hadn’t had so many nail bars around previously,” so you might say that there is a polarisation into high end, high skilled jobs, but also very low skilled personal service sector jobs?
I think that’s absolutely right and, in fact, one bit of analysis we did in our original study was to try and plot the probability of computerisation against two different measures of skill. So, in particular, we considered how the probability of computerisation might relate to the fraction of workers in any occupation who have at least a bachelor’s degree and also the average median wage in each of these occupations.
What we saw, very clearly, is that the probability of computerisation was negatively correlated with these measures of skill, so simply put, the more skilled you are, the less susceptible you are to computerisation, which does indeed suggest only a continuation of this polarisation in the job market.
And a trend that we’ve seen even before that, you know the sort of low skill, personalised service sector jobs; they’re also very hard to replace with technology, aren’t they?
That’s right, but they’re not necessarily going to be very highly paid, because as the result of them not requiring much skill, there’s going to be a whole host of people who are willing to take on these jobs and this, as I suggested, may only lead to an exasperation of the inequality we’ve seen develop over the last couple of decades.
Yes, exactly, what it basically means is that the middle skilled jobs are going to be hollowed out, they’re moving towards both ends, to the very top and the very bottom.
I think that’s correct.
You’ve just mentioned, actually, the next aspect I want to talk to you about, it’s basically inequality. The main point of the paper that I wrote on the issue, at the end of last year, is that we are already back to inequality levels that show certain macroeconomic dysfunctionalities. I mean, you just have to go into Thomas Piketty’s work to look at the fundamentals of this.
But, there is a likelihood with the unfolding impact of technology that these inequalities will get worse, at least in the short to medium term and that we’re going to see an exacerbation of inequality. Would you also see it that way?
I would indeed and, in fact, there is an excellent book that’s just come out from Oxford economist, Tony Atkinson, which puts even more weight behind this hypothesis. I think we need to, however, not lose sight of the fact that there will be enormous wealth generated by technological change.
So, it’s clear that in the development of digital goods that have next to zero marginal cost that innovation has enormous potential of social good for all, at least when we consider people as consumers. But the risk, of course, is that due to these risks to people as workers, there will be developing inequality. I think it’s not necessarily a problem of wealth generation, it’s a problem of distribution.
Fundamentally, yes, it makes the distribution question that we already have even more pressing, in my view. Coming to the policy responses, I agree with everything you’ve said and, in terms of policy response, how should policy-makers react to this or try to prepare for this?
I personally see three areas that need addressing. The first one is the distribution of the remaining work, so there is an argument in the sense that we are back to where John Maynard Keynes was several decades ago, about the need to work longer hours or fewer hours.
So, there might be a question about distributing the remaining work more evenly amongst the population and allowing for more free time, which is basically something that, for instance, German labour market studies tell you is what a lot of people want.
The second would be, especially if you want to prepare for the kind of augmentation role of humans through technology, that the educational profiles need to change. So, educational policy needs to prepare people much more in terms of adaptability. That, to me, means that they need to be better at creative and analytical skills rather than memorising facts, which a lot of the educational systems still seems to rely on.
The third one is really dealing with the social as well as economic consequences of this inequality issue. Is this something that you would broadly see as the categories along which we would have to think and what would you add?
Well, I think you’re right, but I think this problem is an enormously challenging one and I’m not sure there is a simple silver bullet solution. Picking up on some of the suggestions in particular, I think you’re right that education needs to change in response to the accelerating rate of technological change, but how, in particular, it should change, I think is a much thornier issue.
So, you’ve highlighted that, perhaps, our schooling system should emphasise more creativity and put less emphasis upon memorising facts, but there is a question, first, about how you might do that and exactly what kind of creativity should be instilled.
If we take the example of one of the occupations that’s emerged that I mentioned earlier, data scientists, I think actually, to be quite creative as a data scientist, you need to have, at your fingertips, imprinted upon your brain, a large mass of mathematical and statistical facts.
I’m not really clear that you can be creative in that occupation without, at some point, having learnt quite a large amount of knowledge and how that is best to be done, I’m not entirely sure and if indeed that’s the right type of creativity for 20 years hence, I’m not entirely sure. I think it’s very difficult to foresee exactly what skills, in particular, the workforce of 20 years hence will need.
We’re not at a detailed level here yet. At the beginning, these are just to set out, really, the broad compartments, in terms of policy thinking on this. Of course, it’s a very broad brush, but you have to categorise first before you can actually drill down into the details.
I mean, in the distributional question, we have this debate, actually originated from Silicon Valley, about the basic income and its impact on how it can prop up domestic demand in a situation where we have inequality that is going to run away even further. It’s just basically about categorising the policy areas we need to address and then look for individual solutions within.
And fair enough, obviously I endorse that particular effort to categorise possible responses. I’m just concerned that even within these categories, it’s not clear to me that there is something that can actively solve this problem, or that can solve the problem while being, kind of, politically acceptable.
I mean, maybe my pessimism here is borne out of, I mean, I’m an engineer, at the end of the day. I’m not an economist, so the extent to which people are pointing to me as being somehow a source for these policy responses is, to me, already a sign of some sort of desperation.
Okay. But, if you take this away for a moment, your backdrop as an engineer, if you were a policy maker and there were no restrictions to what you could do, what would – against the backdrop of the current knowledge of the subject – your response be?
So, the distributional question that you highlight, I think, is the most important one. With the best educational system in the world, I don’t think we’ll necessarily be able to upskill everyone to the point that they’re able to fill these new jobs that are created. So, then it really does become a question of ensuring that the wealth that is generated is shared equally amongst all members of society.
How that’s to be done, I’m not entirely sure. I think the universal basic income is one kind of exciting possibility. But my background doesn’t leave me best qualified to assess the relative merits of these different proposals.
Okay, well Michael, thank you very much for joining us today on this highly relevant topic and I’m sure it’s not going to go away any time soon.
Well, I hope not, it’s an important one and I hope there will continue to be a lot of discussion around it. Thank you very much for having me.
You can also listen to the podcast using the player below.
We would like to thank the Hans Böckler Stiftung for their generous support for this project.