The Technology Trap

by Carl Benedikt Frey

  Sources: C. B. Frey, T. Berger, and C. Chen, 2018, “Political Machinery: Did Robots Swing the 2016 U.S. Presidential Election?,” Oxford Review of Economic Policy 34 (3): 418–42; W. D. Nordhaus, 2007, “Two Centuries of Productivity Growth in Computing,” Journal of Economic History 67 (1): 128–59; N. Jaimovich and H. E. Siu, 2012, “Job Polarization and Jobless Recoveries” (Working Paper 18334, National Bureau of Economic Research, Cambridge, MA).

  Note: The figure shows how routine employment has contracted as the cost of computing has fallen. All dots denote the year in which a new computer technology was introduced and its cost.

  Focusing only on the rise and fall of individual occupations, however, inevitably glosses over much of the transformation of the workplace. Many of these changes have taken place within occupations. For instance, while the job of the secretary has not disappeared, it no longer has much in common with the jobs held by secretaries in the 1970s. Before the computer revolution, the Bureau of Labor Statistics described the role of the secretary as follows: “Secretaries relieve their employers of routine duties so they can work on more important matters. Although most secretaries type, take shorthand, and deal with callers, the time spent on these duties varies in different types of organizations.”15 The impact of the computer age becomes evident when we look at the description of the same job from the same source in the 2000s: “As technology continues to expand in offices across the Nation, the role of the secretary has greatly evolved. Office automation and organizational restructuring have led secretaries to assume a wide range of new responsibilities once reserved for managerial and professional staff. Many secretaries now provide training and orientation to new staff, conduct research on the internet, and learn to operate new office technologies. In the midst of these changes, however, their core responsibilities have remained much the same—performing and coordinating an office’s administrative activities and ensuring that information is disseminated to staff and clients.”16 What is true of secretarial positions is also true of many other jobs. For example, in the 1970s, American men and women could make a good living as bank tellers by accepting deposits and paying out withdrawals. As noted, the job has not disappeared, but the skill requirements have changed so dramatically that it requires a different breed of worker.

  All the same, computerization has clearly not been as dismal for labor as some people predicted when ENIAC arrived. Although computers have taken over an ever-growing share of routine work, labor has retained its comparative advantage in other domains. One reason is because of what the economist David Autor has called “Polanyi’s paradox.”17 A key bottleneck to automation that engineers have found hard to overcome is well summarized by Michael Polanyi’s famous phrase: “We know more than we can tell.”18 (We shall take a closer look at the AI-enabled inroads on Polanyi’s paradox in chapter 12.) Humans constantly draw upon large reservoirs of tacit knowledge that we struggle to articulate and define even to ourselves, making it exceedingly hard to specify it in computer code. To illustrate Polanyi’s point, it is helpful to contrast the task of repetitive assembly with that of designing a new car, writing a piece of music, or giving a galvanizing speech. The rules for what makes a good song or a great speech are hard to define because there are none. Artists and other creative professionals constantly break and redefine them. From an automation point of view, Polanyi’s insight is critical, because it means that there are many tasks humans are able to perform intuitively but that are hard to automate because we struggle to define rules that describe them. For activities that demand creative thinking, problem solving, judgment, and common sense, we understand the skills only tacitly. But more importantly, from an economic standpoint, Polanyi’s observation also means that some human skills are complemented by computers. Things done by computer technology, such as the storing and processing of information, make humans more productive problem solvers, decision makers, and analysts. As computerization reduced the costs of critical inputs to these tasks, humans have become more productive in computer-using jobs.19 In 1970, a lawyer in Grinnell, Iowa (a town without a law library), would have had to drive to the next city to do legal research. In the computer era, she can leave her car in the garage and connect her word processor to Westlaw, which provides digitized records on case law, state and federal statutes, administrative codes, and so on. Indeed, most of the professions have found their practitioners’ skills augmented by computers. Consider, for example, the evolving office of Stephen Saltz, a Boston cardiologist:

  In September 2001, Dr. Saltz took an echocardiogram of an elderly male patient we will call Harold. Harold had suffered a small heart attack. His condition was complicated by diabetes, a disease that creates “silent” heart blockages not detected in standard tests. When Saltz had trained in Boston’s Brigham Hospital in the early 1970s, an echocardiograph was an oscilloscope-like device that provided limited information on the heart’s blood flow and valve flaps. Over time, advances in computerization allowed the instrument to create a full two-dimensional image of virtually all aspects of the heart’s functioning, including blood flows, blockages, and valve leakages. Using this image, Saltz saw that the entire front wall of Harold’s heart was malfunctioning. The information led Saltz to refer Harold to a surgeon who would perform a bypass or insert a stent. Either operation would improve the length and quality of Harold’s life. The computerized imaging had made Saltz a better diagnostician.20

  As engineers have expanded what computers can do, technological progress has continuously moved in the direction of favoring skills that require higher education, such as complex problem solving and creative thinking, because computers have taken over the more mundane tasks. When Robert Reich surveyed the transformation of the labor market in his classic 1991 book, The Work of Nations, he found that work could be divided into three broad categories. A new class of what he called “symbolic analysts” had emerged, who were reaping the benefits of the new economy.21 Among these analysts, we find managers, engineers, attorneys, scientists, journalists, consultants, and other knowledge workers. In the age of computers, they had all become more productive analysts. Besides symbol-analytic services, Reich reckoned, there are also routine jobs and in-person services. As noted above, routine jobs have gradually been taken over by computers. But in-person service jobs have become more plentiful. Indeed, most Americans do not work in technology industries or professional services. Few are employed directly by software companies, law firms, or biotechnology start-ups. But these occupations still support the livelihoods of many citizens. While today’s technology companies provide less opportunity for the unskilled relative to the smokestack industries of the Second Industrial Revolution, many people are employed indirectly by those firms, whose employees demand many services that the unskilled provide. When America’s symbolic analysts shop locally, they support the incomes of hairstylists, barkeepers, waiters, taxi drivers, and store clerks. These jobs may not have seen the technological miracles of biotechnology and software production, “but that is where most Americans work and their fates are tied to their ongoing ability to sell their time to those workers who sell exportable goods and services.”22

  If Polanyi’s paradox were the only hurdle to automation, most remaining jobs would be for symbolic analysts. A second reason why there are still so many jobs is explained by Moravec’s paradox, named after the computer scientist Hans Moravec. The paradox he noted was the fact that it is hard for computers to do many tasks that are easy for humans, and conversely, computers can do many things that we find exceedingly difficult: “It is comparatively easy to make computers exhibit adult level performance in solving problems on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.”23 A computer would have an easy time beating the world chess champion Magnus Carlsen, but it would be unable to clean the chess pieces after the game and put them back in the right place. Any human cleaner still outperforms computer
-controlled machines in perception, dexterity, and mobility. Today’s computers far exceed human capabilities in storing and processing information, yet they cannot climb a tree, open a door, clear coffee cups off a table, or play football. A powerful explanation is that our unconscious sensorimotor abilities have evolved in the human brain over millions of years, making them exceedingly difficult to imitate. From an early age, humans can walk, identify and manipulate objects, and understand complex language. Giving a computer these basic abilities, mastered by any four-year-old child, has turned out to be among the hardest engineering problems.

  Relative to Polanyi’s paradox, the critical difference is that many of the skills that are hard to automate because of Moravec’s paradox have not been made more valuable by computers. Together, the persistence of these engineering bottlenecks explains why the labor market has evolved the way it has. As computers have made symbolic analysts more affluent, the analysts have spent a larger percentage of their income on personal services that are hard to automate. But the automation of routine jobs has meant fewer employment opportunities for high school graduates, so there has been a flow of workers from productive and automated sectors to low-productivity service jobs—like those of janitors and gardeners, child-care workers, receptionists, and so on.24 Unfortunately, this has meant that millions of workers have migrated into jobs where the productivity ceiling is low, and consequently, their wages have fallen behind those of symbolic analysts. Even so, economists would expect the wages of people in technologically stagnant jobs to rise, as employers need to pay them more to prevent workers from leaving for more productive jobs of higher pay. As noted, the fact that the wages of men without college degrees have fallen over the course of three decades, in other words, suggest that they are faced with fewer alternative job options for which their skills are suitable. Together with Autor, in their pioneering 2004 book, The New Division of Labor, Frank Levy and Richard Murnane, two economists at the Massachusetts Institute of Technology, were among the first to note this pattern:

  As computers have helped channel economic growth, two quite different types of jobs have increased in number, jobs that pay very different wages. Jobs held by the working poor—janitors, cafeteria workers, security guards—have grown in relative importance. But the greater job growth has taken place in the upper part of the pay distribution—managers, doctors, lawyers, engineers, teachers, technicians. Three facts about these latter jobs stand out: they pay well, they require extensive skills, and most people in these jobs rely on computers to increase their productivity. This hollowing-out of the occupational structure—more janitors and more managers—is heavily influenced by the computerization of work.25

  While their work focused on the United States, the polarization they observed was not just an American phenomenon. As figure 12 shows, the hollowing out of the middle is a feature of labor markets across the industrial world. This dynamic of job growth at the top and the bottom of the skill and income distribution has contributed to the growing divide between college and high school graduates.

  FIGURE 12: Job Polarization in 16 European Countries, 1993–2010

  Source: M. Goos, A. Manning, and A. Salomons, 2014, “Explaining Job Polarization: Routine-Biased Technological Change and Offshoring,” American Economic Review 104 (8): 2509–26.

  The Cognitive Divide

  Terms like “middle class” and “working class” were invented to describe the profound changes that accompanied the Industrial Revolution. Yet as of late, they have become increasingly problematic. President Barack Obama mentioned the “middle class” ten times in his first State of the Union address. He spoke about the “working class” only once, when he referred to Vice President Joseph Biden as “a working-class kid from Scranton.”26 The disappearance of industrial jobs has meant that fewer citizens can be regarded as working class. Most adults with a high school education no longer work in factories. There is no stable working class for young men and women to break into. Thus, the term has become a shunned pejorative, while the “middle class” is now used to refer to almost everyone except the very wealthy and the very poor.27

  What we regard as “middle class” has, of course, always been elastic. During the classic years of industrialization, it was largely reserved for the commercial and industrial bourgeoisie. But in the succeeding centuries, it expanded to the point that it converged with the working class. Welding-machine operators and other blue-collar workers became associated with relatively stable, decent-paying jobs by the mid-twentieth century. The conditions of the working classes in the golden postwar years were strikingly different to those described by Karl Marx and Friedrich Engels a century earlier, and they were able to attain a middle-class lifestyle. But in a world of robots, there are fewer jobs for the so-called blue collar aristocracy. With some rare exceptions, only college-educated workers qualify as middle class. What distinguishes Reich’s symbolic analysts from the rest is that nearly all the jobs they hold require a college degree. As clerical and blue-collar jobs have disappeared from the lower and middle parts of the income distribution, the employment prospects for young people with no more than a high school education have become more similar to those of high school dropouts than to people with a college education. Thus, sociologists mostly use college education, rather than occupation, as an indicator of a citizen’s class in the post-1980 period.28

  As has been widely documented, education has reinforced the divide between those who thrive in the new economy and their less-educated peers. This pattern becomes all the more evident when we look at how workers have adjusted to automation. Those with analytical skills have moved up into the expanding sets of high-wage jobs, while people who lack valuable skills have dropped down and are competing for unskilled service jobs at declining wages. In the postwar era, workmen on the assembly lines who experienced displacement could still find work in other routine jobs requiring similar skills. But since the computer revolution, unemployed Americans who used to work in a routine occupation have become much less likely to find new employment in routine jobs.29 Fewer job options, especially for non-college-educated production workers, has led to cascading competition for low-skilled jobs.

  To be sure, as the jobs of machine operators dried up, new highly skilled jobs were created, as computer programmers were needed to design numerically controlled machine tools. In 1985, when the first wave of automation swept through America’s car factories, the Wall Street Journal published a story about Lawrence Maczuga, a thirty-seven-year-old machine operator at Ford Motor Company’s transmission plant in Livonia, Michigan. While Maczuga was working the machines at the factory, he had been going to college at night to get a degree in computer science. As automation took over, he gave up his semiskilled job on the assembly line to take one of the plant’s “superjobs” as a manufacturing technician. Maczuga had for some time considered giving up his job at Ford to become a computer programmer. But as Ford revealed its new plans for computerizing production, he accepted the new role.30

  The problem is that Maczuga was the exception, not the pattern. Few manual workers got degrees in computer science or any other college concentration. Thus, as automation eroded the demand for routine skills and physical strength, blue-collar people found themselves in an ever-weaker position. Indeed, the economists Matias Cortes, Nir Jaimovich, and Henry Siu found that prime-aged men without college degrees were the main victims of the contraction of routine employment. Many adjusted by taking on low-paying service jobs, like those of food-preparation workers, gardeners, and security guards. Unskilled men were more likely to find themselves pushed down or even out of the workforce as routine jobs dried up than they were to move up.31

  The adverse consequences of automation have manifested themselves not only in falling wages but also in rising joblessness among groups in the labor market. For several decades now, the percentage of prime-aged men ages 25–54 who do not go to work in the morning has steadily risen (figure 13). While economists still debate
the relative importance of supply and demand factors in explaining men’s detachment from the labor market, there is an emerging consensus that demand factors should be given more weight for recent years. Welfare programs, spousal employment, and changing social norms have all played a role in shaping some men’s decision not to work over the course of the twentieth century (see also figure 7). But since 2000, most of the rise in joblessness is seemingly involuntary. When the economists Katharine Abraham and Melissa Kearney recently set out to review what we know about the causes of male joblessness, they found trade and robots to be the prime reasons why fewer men ages 25–54 have been working since the new millennia.32

  FIGURE 13: Labor Force Participation Rate (Ages 25–54) by Educational Attainment, 1976–2016

  Source: 1963–91: Current Population Survey data from D. Acemoglu and D. H. Autor, 2011, “Skills, Tasks and Technologies: Implications for Employment and Earnings,” in Handbook of Labor Economics, ed. David Card and Orley Ashenfelter, 4:1043–171 (Amsterdam: Elsevier). 1992–2017: author’s analysis using data from S. Ruggles et al., 2018, IPUMS USA, version 8.0 (dataset), https://usa.ipums.org/usa/.

  However, the experience of women has been rather different. As is well known, the great leap forward of the “pink-collar” workforce came to an end in the 2000s, when computers began to take over more clerical work (figure 13). Just a few decades ago, people who called Amtrak to make a train reservation would have heard a woman answering the phone on the other end. Today, they hear a recorded voice saying, “Hi, I’m Julie, Amtrak’s automated agent.” But consistent with what we know from studies in neuroscience pointing out that women perform better in interactive and social settings, women have adjusted much better than men to an increasingly interactive world of work.33 Instead of being pushed back into low-wage service jobs, where women had traditionally been dominant, many have moved up into professional and managerial jobs. Women also are more likely to graduate from college than men, and consequently their skills are more suitable for the computer age. Indeed, while men have found themselves increasingly likely to be replaced by computer-controlled machines, women are more likely to use a computer at work.34 The rising share of women in the professions and the decline of male-dominated blue-collar sectors have allowed many women to overtake their male counterparts in terms of career advancement. Of course, women still have some way to go before they surpass men in terms of earnings, but a shift is under way. American women age thirty or younger have higher earnings power than their male counterparts—with the exception of the three largest metropolitan areas, where skilled men have clustered.