Chapter 1. The challenge computers pose to work and education

Past employment trends in skill demand

The transformation of work skills has been a key aspect of global economic development over the past centuries. The nature of this transformation is well known, involving the long-term shift of employment out of agriculture into manufacturing initially, and then into services. The shift is accompanied by large increases in educational attainment. It is helpful to remember the scale of this transformation: in the United Kingdom for example, employment in services increased from 41% of the workforce in 1890 to 72% in 1998 (Maddison, 2003, Table 2-24), while average educational attainment increased from 4.8 years to 13.1 years (Van Zanden et al., 2014, Table 5.5). Different countries are at different stages in this transformation of education and work skills, but the transformation is occurring worldwide and has continuing implications for government policies related to human capital.

The overall shift in employment and increasing demand for education are related to technological change, with new technology during the 20th century tending on average to increase the demand for higher skills and decrease the demand for lower skills. This basic historical relationship between technology and education suggests the metaphor of a “race” between technology and education (Goldin and Katz, 2008; Tinbergen, 1974). The general conclusion that technological change is driving the economy towards ever-increasing demands for education is widely accepted. However, the metaphor does not apply to the 19th century, when the most salient effect of technological change was to decrease the demand for skills by replacing skilled craft workers with unskilled workers in factories (Acemoglu, 2002).

In the late 20th century, the change in skill demand became more complex, with the emergence for several decades of a pattern of “polarisation”. This term refers to increasing employment for workers with higher and lower skills and decreasing employment for workers with mid-level skills. This pattern of job change has been found for the United States, many European countries and Japan (Autor, Katz and Kearney, 2006; Goos, Manning and Salomons, 2009; Ikenaga and Kambayashi, 2010). However, these trends will not necessarily continue. Already, there are questions about whether the pattern of polarised employment changes has continued in recent years, especially with respect to findings of weakness in the demand for workers with higher skills since 2000 (Autor, 2015; Beaudry, Green and Sand, 2016).

Although there are several possible reasons for the trend towards polarisation in the labour market, the strongest explanation is that technology is increasingly being used to perform routine tasks that were previously performed by workers (Goos, Manning and Salomons, 2014; Michaels, Natraj and Van Reenen, 2014). According to this explanation, jobs involving routine cognitive tasks typically occur in the middle of the skill distribution and are susceptible to substitution by technology. On the other hand, jobs involving non-routine tasks that cannot yet be carried out by technology occur either at the low or high end of the skill distribution, depending on whether the non-routine tasks require physical or cognitive skills (Autor, Levy and Murnane, 2003).

In general, the research looking at skill polarisation uses indirect measures that are available in economic datasets, with nothing like the skill assessments that are used in education research. Typically, “skill” is inferred indirectly from wages, education or occupation. For example, Autor, Katz and Kearney (2006) use three different measures of skill by occupation: mean level of education by occupation; median level of hourly wages by occupation; and data on tasks taken from brief occupational descriptions. Such measures can be useful as rough indicators of skill in the context of economic analyses. However they are far removed from direct measures of skill.

As technology advances, automation of certain skills is raising questions about the changes in the quantity of jobs that will be needed in the future, not just the skill distribution required across the workforce. Historically, there have been periodic waves of concern about unemployment resulting from new technology. This stretches back at least two centuries to the early industrial revolution (Mokyr, Vickers and Ziebarth, 2015). However, unemployment resulting from technological displacement in the past has always been temporary, with increased productivity leading to decreased prices, which lead in turn to increased demand by consumers for both old and new goods, and then to increased demand by employers for workers (Autor, 2015). This adjustment process may not necessarily take place as smoothly as suggested by economic theory (Autor, Dorn and Hanson, 2016), leaving room for policy interventions to help move displaced workers to new occupations. However, the overall historical experience is one of large-scale and successful redeployment of the workforce as technology shifts the skills needed in different production processes.

In response to several high-profile studies suggesting the possibility of substantial job displacement resulting from computers (Brynjolfsson and McAfee, 2014; Frey and Osborne, 2013), a number of economists have recently looked directly at the employment implications of these technologies. Several studies have specifically looked at the impact of recent applications of computers on employment resulting from applications of computers (Bessen, 2015; Falk and Biagi, 2015; Graetz and Michaels, 2015). In general, these studies have found familiar conclusions about the job effects of this technology. In short, firms, occupations and industries that use higher levels of computers in production experience higher productivity and employment. On the other hand, a recent study looking at the effects of trade and offshoring on jobs in the United States found that it was computer use in occupations, not trade or offshoring that led to increased risk of unemployment (Ebenstein, Harrison and McMillan, 2015).

Projected future trends in skill demand

Looking forwards, several recent economic studies have developed theoretical frameworks to explore the potential effects of computers on employment, wages and productivity in the future. These studies extend earlier theoretical work from a period when the technology was substantially less advanced (e.g., Elliott, 1998; Simon, 1977; Zeira, 1998). Acemoglu and Restrepo (2016) develop a model where automation can displace workers in performing older tasks, but where there is endless creation of new and more complex tasks that workers can perform better than machines. Benzell, Kotlikoff, Lagarda and Sachs (2015) develop a model where computers can automate analytical tasks, but not empathetic or interpersonal tasks. Sachs and colleagues (2015) compare a two-task model, where computers can automate only one of the tasks, with a one-task model where the only task in the economy can be performed by both people and machines. These different models project a variety of results for workers, some positive and some negative, with the results depending on assumptions about the fundamental relationship between computer capabilities and the skills needed to perform tasks in the economy.

Several studies have used information from computer science to understand the relationship between computer capabilities and the skills and tasks in the economy. The most widely cited study is by Frey and Osborne (2013), which estimates that 47% of employment in the United States is at a high risk of automation over the next several decades. The Frey and Osborne analysis involves four steps. First, a group of computer scientists classified 70 occupations as either automatable or not, based on a set of job descriptions and their knowledge of current computer capabilities.4 The occupations chosen were those where the group was most confident in making this judgment. Second, the authors identified job tasks that were most likely to be barriers to computerisation – perception and manipulation, creative intelligence, social intelligence – based on the current state of computer science. Third, occupational data on these hard-to-automate tasks were used to develop a model to predict the “automatability” classification of the 70 occupations from their tasks. This model was then used to predict the automatability of all occupations in the United States economy. Finally, jobs above a predicted automatability of 70% were defined as “high risk”. Figures of employment by occupation were then used to derive the overall estimate of 47%. Although the model was originally developed for the United States, it has been applied to a variety of other countries by substituting different figures for occupational employment in the last step (e.g., Frey, Osborne and Holmes, 2016; Pajarinen, Rouvinen, and Ekeland, 2015).

Further work supported by the OECD has developed an alternative way of extending Frey and Osborne’s automatability judgments for the original 70 occupations to the full economy (Arntz, Gregory and Zierahn, 2016). This research suggests an estimated 9% of jobs in OECD countries are highly automatable, a dramatically lower figure from that estimated by Frey and Osborne from the same starting point. Unlike Frey and Osborne, who use job tasks that are hard to automate as the basis for extending the automatability rating beyond the original 70 occupations, the analysis by Arntz, Gregory and Zierahn uses information on a wide range of job tasks, job characteristics, worker skills and worker characteristics from the OECD’s Survey of Adult Skills (PIAAC). This information includes factors that are not directly related to job tasks such as gender, age, education, proficiency in literacy and numeracy, firm size and income. In addition, the job task information includes not only factors that are similar to the hard-to-automate tasks identified by Frey and Osborne, but also other tasks that may not be hard to automate, such as filling forms or calculating percentages. Arntz and colleagues attribute their differing results to a “task-based approach” that acknowledges the variation of job tasks within an occupation, using the job task information in PIAAC. However, they do not provide any analyses to prove that it is the variation within occupation by tasks that explains the substantial difference in their results from those of Frey and Osborne. Another plausible explanation of their result is that it comes from using a very different set of job features that includes many things that are not job tasks at all. Their report does not include any results for models that use job tasks alone as the basis for the extrapolation.

The McKinsey Global Institute recently issued a report analysing work automatability using an approach that focuses on judgments related to 18 capabilities that are mapped to more than 2 000 work activities in the United States and other countries (Manyika et al., 2017). The report estimates that 49% of the activities at work could be automated with current technology. The report does not describe how the judgments related to the 18 capabilities were obtained. The capability judgments were mapped to work activities by a process using the key words in the titles of the work activities.

Another approach to predicting job automatability focuses on worker skills rather than job tasks or activities. Although these two approaches should be complementary, a focus on skills may be more meaningful to the education community. Elliott (2017) uses a sample of recent articles from the computer science literature to identify computer capabilities in four rough skill areas – language, reasoning, movement and vision – that can be mapped to a set of worker descriptors in occupational data for the United States (O*NET). The descriptions of computer capabilities from the research literature are compared to the anchoring tasks on the O*NET scales. The resulting analysis suggests that 82% of current United States employment is potentially automatable, based upon the types of capabilities discussed in contemporary computer science literature.

All of these studies making projections about the potential automatability of current jobs using current technology include caveats about the various economic, institutional and social factors that affect the application of technology. Extensive literature explores the factors that influence the diffusion of innovation (Rogers, 1995). When diffusion does occur it can often take several decades or more (Comin and Hobijn, 2010; Griliches, 1957; Mansfield, 1961), even though diffusion speed has increased in recent years. None of the studies of potential automatability focus on an exact timeframe. All of them refer loosely to computer applications that could potentially happen over a period of several decades. Of course, it is quite possible that some potential applications of current technology will not occur over this period. It is also possible that even more advanced technology will be developed. All of the studies note that projection of automatability of a percentage of jobs, occupations or tasks over several decades does not mean that the people who currently perform those activities will become unemployed, or even that they will change jobs. Instead, the crucial question relates to the ways in which their activities and required skills will be redefined, whether or not their job or occupation changes over this period.

Separately within computer science, there is also work to assess the capabilities of current computer techniques with standardised tests developed for people. Early versions of this work go back several decades (O’Neil and Baker, 1994). This research encompasses several different types of tests, including elementary and secondary school tests in science and mathematics (Clark and Etzioni, 2016), verbal IQ tests for young children (Ohlsson et al., 2015) and university entrance exams (Arai and Matsuzaki, 2014). To work towards the goal of a broader development and assessment of artificial intelligence, computer scientists have proposed formal assessments in social-emotional intelligence (Jarrold and Yeh, 2016), physical perception and action (Ortiz, 2016), visual interpretation (Zitnick et al., 2016) and common sense reasoning (Davis, 2016).

Plan for the study and report

This study builds on this prior research related to past and future trends in work skills by using the OECD’s Survey of Adult Skills (PIAAC) to look at the use of general cognitive skills in the workplace. Chapter 2 looks to history, using PIAAC to describe the distribution of proficiency in the workforce, the use of skill and the ways these have changed over the past two decades. Chapters 3 and 4 outline the approach of using PIAAC to measure computer capabilities against workforce skills, with Chapter 3 describing the development and Chapter 4 the results. These two chapters provide a way of understanding a technology that could change skill demand in the decades ahead. Chapter 5 discusses the implications of computer capabilities for the future of the skill changes discussed in Chapter 2 and considers the policy implications for education.