Thank you very much. It's an honor to be here. And actually, let me put up my slides. So my presentation is called the shaping the future work.
And actually this builds on comments, both from Mike Spence and Erik Berglöf. I think when we think about predicting, thinking about the future and anticipating the future, we often take a passive view of the future will occur in our role is to determine how we're going to respond to it. But I want to emphasize the point in this very brief remarks. Actually we have a lot of agency in shaping the future that we are going to experience. And Erik Berglöf also hinted this in his remarks that new technologies opened new possibilities. Not only the opening of possibilities, but the technologies themselves are things that we create and influence. There's a grave concern you all are well aware of that automation is rendering human labor superfluous. Firstly articulated by Daniel Susskind in 2020: “Machines will not do everything in future, but they do more. And as they slowly but relentlessly take on more and more tasks, human beings will be forced to retreat to an ever-shrinking set of activities.” This is not a lump-of-labor fallacy argument. This is sophisticated argument about comparative vantage. Wassily Leontief said roughly the same thing in 1983, when he said humans will be rendered to the position of horses in that they are no longer a competitive factor of supply for many tasks.
We know that after two centuries of rapid automation, we're still hard at work. The employment to population ratio in most advanced countries and many developing countries has been rising, mostly as women have moved out of unpaid, unremunerated, very restrictive labor in the household to market opportunities and skills.
So there are three reasons why there are still so many jobs. Two of them, I think are quite familiar. One is insatiability. As we get wealthier, our consumption desires scale at rate about 1.03 times the rate of increase of our incomes. And that creates a lot of demand and new employment. The second is complementarity. Automation or technology doesn't simply replace us. It often makes us more effective at what we do, whether you can use a pneumatic nail gun to install a roof or whether you're using a statistics package to do data analysis. Complementarity often takes our existing skills and magnifies their reach, and so increases our marginal productivity.
For the third point, and the focus of my remarks today is about new work. Novel demands for human specialization, expertise that are enable brought about by changes in technology, but also other demographic and social changes that create new needs for our specialized skills or new skills that didn't previously exist. So in 1940, US employment was about 45 million workers. In 2018, if we look at employment in 2018 in job titles, and I'll define my terms in a minute, that existed in 1940, employment hadn't actually risen that much risen by about 5 million workers, but we know the US labor force tripled in that time. So what are people doing?
Well about, 2/3 of employment in 2018 is in detailed jobs that were not been present or not yet been invented in 1940. Let me kind of give you a like a very illustrative example. What is new work. So here's an example of marketing research. So timeless task in marketing research: form hypotheses, propose analysis, stare at results, start again. Many new tasks are done by machines: estimating models, testing significance, generating tables and figures. Simultaneously, new work has been created in this occupation. So in 1980, the job applied statisticians entered into the US census and also director of marketing research analysis. In 2018, data visualization developer. These are not just new titles. These involve new specializations, new skill sets that weren't demanded in prior decades.
Or take another example, fitness training. They are timeless tasks: grunting, and sweating the agony and, the ecstasy, et cetera. Their tasks are formerly done by workers now done by machines: time keeping, heart rate, pace, adjusting resistance, and weights on equipment. There are also new jobs: a sports psychologist, sports nutritionist, certified therapeutic recreation specialist. A lot of new specialization that require scare skills. Broadly, looking across the sweep of the last 80 years, I've now arranged all US occupations into 12 broad categories, roughly ordered from low paid to high paid. You can see that in 1940, that's the blue bars about 40% of employment was just in two categories: in farming & mining and in production. If you look in 2018, you'll see that actually the number of workers, not the share of workers, the number of workers in farming & mining and production has fallen despite the tripling of the size of labor force. There's much more employment now in professional jobs, managerial jobs, in sales jobs, clerical administrative support jobs, and personal service jobs. And in many cases, the red bar, most of that work is new work that was not present in decades earlier.
There are four points I’m going to make and I'm going to run out of time so let me just summarize my points. One is that technological innovations both augment and automate human labor. These are distinct processes. We can measure them, and we can distinguish them empirically. It's not just technology. Demand forces also shape where new work appears. Negative trade shocks, for example, discourage new work creation. I don't mean new employment creation. I mean the introduction of new types of specialized work. Demographic pressures conversely can foster new work creation, both these things can work in both directions. It matters, therefore, not just whether we innovate, but how and where we innovate. How do we augment or do we automate? Where and which occupations are augmented, which ones were automated? Finally, that implies that the future of work is a work in progress. The jobs we get dependent part on the investments that we make and the institutions that turn productivity into prosperity. I should say my collaborators on this work on the empirical part are Anna Salomons at Utrecht University and Bryan Seegmiller, who's a PhD student at the MIT Sloan School. So in the remaining five and a half minutes, I will say something about how we measure new work, several hypotheses we test and then a couple conclusions we draw.
First, in terms of measuring. How do we measure new work? We use historical census documents that basically, when you fill in the census, you have to write in your occupation. It's not a check box or bubble list. You write in what you do. The Census Bureau then has to code those into standardized categories in which they're between 300 to 600 per decade. The way it does this is to create a phone book of things you might write in will be 30,000 to 40,000 examples in each decade. It uses those to classify. But when a new example starts appearing, the Census hasn't seen before it adds them to these volumes. These volumes capture the flow of new titles as they enter usage.
For example, in 1940, the job of automatic welding machine operator was added. In 1960, textile chemist. In 1980, controllers for remotely piloted vehicles. In 2000, AI specialist. In 2018, pediatric vascular surgeons. So you might look at this and say I get it. New work is about using new technology, designing new technology, installing new technology, selling new technology, integrating new technology and maintaining new technology. Now, let me give you another set of examples, also added to the census: acrobatic dancers in 1940, pageant directors in 1960 hypnotherapist in 1980, sommelier in 2010, drama therapist in 2018. Many of these things do not have a technological component necessarily. They reflect rising incomes, changing tastes, changing demographics and the change the need for child care or adult care.
So broadly, new work is not just about new machines. It's often about new specialties involving things assisting others or entertaining others or serving others. Just so you know, how do I know these titles are new? These are the ones you pulled out a Census volume. Here I’m using the Google Ngram Viewer that measures the frequency of words occurrences in the US in published US texts. And I'm looking at the ratio of occurrence of these new titles to existing titles. You can see here titles that were added in 1940, they have a peak that we find in 1940 and ones in 1950, 60, 70, and so on. So you can see that their appearance in these Census volumes corresponds reasonably well with their frequency of occurrence of usage in published US language texts. So we think these were measuring something real here.
We try in this work to measure augmentation and automation. By augmentation, I mean technologies that add value to worker tasks. How do we do that? We again use that Census volumes that describe these detailed titles. And we think of these as describing the output of occupations, for example, and anesthesiologist assistant, or an audiometrist, or an orthopedic mechanic, or an optometric technologist. So we think of those as describing occupational outputs, we're going to look for innovations that discuss those outputs. We're going to think of those innovations as potentially augmenting the value, creating changing demand, or changing the set of activities that can be accomplished.
We also want to look for automation innovations, things that substitute for worker tasks. The way we're going to define those is we're going to use the Dictionary of Occupational Titles that describes workers task input into occupations. We’re going to look for innovations that claim to do what workers do, whether that is performing blood tests or analyzing test results, or providing data for diagnosis, et cetera.
So we’ll try to distinguish those two things. I'll say in a sentence what we do. We actually use natural language processing. We take both these descriptions of occupational outputs, these descriptions of occupational inputs, and all US patents from 1930 to present. We turn them into word embeddings. We find overlaps effectively using cosine similarity measures. We get the count of highly cited patents that appear to augmented occupation and those that appear to automate an occupation. That's our basic approach. It's very hands-off. This is not cherry-picked. This is not tuned. It's completely a parable for augmentation and automation. We show that in general, occupations that are exposed to automation are also exposed to augmentation. Here's automation. Here's augmentation. For example, assemblers of electrical equipment are exposed to a lot of both. Child care workers aren't exposed to much of either. However, there are some occupations that appear to be being automated, not augmented: machinist, cabinet makers, elevator installers. Others that appear to get a lot of augmentation like operations and systems research analysts, or business and promotion agents.
There's then three hypotheses that we test. I'm just going to tell you what they are, and I won't tell you how we test them. One is that augmentation technologies spur the creation of new job types. Where we see augmentation occurring, we see new job titles emerging. That's not true for automation technologies. Those do not predict the emergence of new job titles. Two, we show that demand shifts, not just technologies, spur and retard the creation of new job types. When you have an outward demand shift, for example a demographic shift, not only is it creates more jobs, it creates more specialized work. When the China trade shock causes a big decline in US manufacturing employment due to China's rising productivity, we see not only a decline in manufacturing workers, but a slow down in the rate of new manufacturing specialization. Finally, this is not just about new job titles but about aggregate changes in work, where we see a lot of new titles emerging, we see employment growing where there's augmentation. Where we see a lot of automation, we see employment contracting.
Let me just skip all that and just say, what do we draw from that? The first key take away is the jobs we get depend in part on the investments we make. This figure shows you the changing shape of innovation. As we move from kind of a machine age to a chemical age, finally to an information age, where more than 50% of all patents are information & instruments and electricity & electronics. That changes the set of occupations exposed to innovation. And I’ll just summarize them as I’m out of time. It dramatically changes where new work appears. In the early postwar era, most new work for non-college workers was in middle paid construction, transportation, production and clerical. It’s now mostly in low paid personal services. Most new job titles emerging for high skilled workers are increasingly specialized. We see this even during the pandemic that we see augmentation technologies and innovations being directed at work from home, which will augment the productivity of high skill workers going forward.
So I’m going to summarize. New work is quantitatively important. More than 60 % of 2018 employment is in job titles that didn't yet exist in 1940. Innovations do predict where new jobs emerge. New titles emerge where innovation compliments outputs and they do not emerge where innovation replaces inputs. Innovation is only part of the story. New work is driven by changing taste, wealth, demographics, globalization. Its emergence accelerates and decelerates in predictable ways.
So the conclusion that I'm going to draw is that the work of the future is a work in progress as I said. The majority of today's jobs have yet to be invented. We are inventing the work of the future right now. We shape what technologies, whether they augment, whether they automate and whom. So we should, knowing that agency we have, we should invent a future that we would want our children to inherit. Thanks very much.
For more information, please visit Luohan Academy's youtube channel: Luohan Academy