+1(978)310-4246 credencewriters@gmail.com
  

Description

Technological Unemployment

This week I’d like you to think about the automation of labor and its impact on the future of jobs and the changing economy. Continuing our conversation about artificial intelligence, consider how automation might also change our attitudes towards work, value, and our relationships to each other.

sources:-

https://books.google.com/books?id=cccDAAAAMBAJ&lpg…

THE FUTURE OF EMPLOYMENT: HOW
SUSCEPTIBLE ARE JOBS TO
COMPUTERISATION?∗
Carl Benedikt Frey† and Michael A. Osborne‡
September 17, 2013
.
Abstract
We examine how susceptible jobs are to computerisation. To assess this, we begin by implementing a novel methodology to estimate
the probability of computerisation for 702 detailed occupations, using a
Gaussian process classifier. Based on these estimates, we examine expected impacts of future computerisation on US labour market outcomes,
with the primary objective of analysing the number of jobs at risk and
the relationship between an occupation’s probability of computerisation,
wages and educational attainment. According to our estimates, about 47
percent of total
US
employment is at risk. We further provide evidence
that wages and educational attainment exhibit a strong negative relationship with an occupation’s probability of computerisation.
Keywords: Occupational Choice, Technological Change, Wage Inequality, Employment, Skill Demand
JEL
Classification: E24, J24, J31, J62, O33.
We thank the Oxford University Engineering Sciences Department and the Oxford Martin Programme on the Impacts of Future Technology for hosting the “Machines and Employment” Workshop. We are indebted to Stuart Armstrong, Nick Bostrom, Eris Chinellato, Mark
Cummins, Daniel Dewey, David Dorn, Alex Flint, Claudia Goldin, John Muellbauer, Vincent
Mueller, Paul Newman, Seán Ó hÉigeartaigh, Anders Sandberg, Murray Shanahan, and Keith
Woolcock for their excellent suggestions.
†
Oxford Martin School, University of Oxford, Oxford, OX1 1PT, United Kingdom,
carl.frey@oxfordmartin.ox.ac.uk.
‡
Department of Engineering Science, University of Oxford, Oxford, OX1 3PJ, United Kingdom, mosb@robots.ox.ac.uk.
∗
1
I.
I NTRODUCTION
In this paper, we address the question: how susceptible are jobs to computerisation? Doing so, we build on the existing literature in two ways. First, drawing
upon recent advances in Machine Learning (ML) and Mobile Robotics (MR),
we develop a novel methodology to categorise occupations according to their
susceptibility to computerisation.1 Second, we implement this methodology to
estimate the probability of computerisation for 702 detailed occupations, and
examine expected impacts of future computerisation on US labour market outcomes.
Our paper is motivated by John Maynard Keynes’s frequently cited prediction of widespread technological unemployment “due to our discovery of
means of economising the use of labour outrunning the pace at which we
can find new uses for labour” (Keynes, 1933, p. 3). Indeed, over the past
decades, computers have substituted for a number of jobs, including the functions of bookkeepers, cashiers and telephone operators (Bresnahan, 1999; MGI,
2013). More recently, the poor performance of labour markets across advanced
economies has intensified the debate about technological unemployment among
economists. While there is ongoing disagreement about the driving forces
behind the persistently high unemployment rates, a number of scholars have
pointed at computer-controlled equipment as a possible explanation for recent
jobless growth (see, for example, Brynjolfsson and McAfee, 2011).2
The impact of computerisation on labour market outcomes is well-established
in the literature, documenting the decline of employment in routine intensive
occupations – i.e. occupations mainly consisting of tasks following well-defined
procedures that can easily be performed by sophisticated algorithms. For example, studies by Charles, et al. (2013) and Jaimovich and Siu (2012) emphasise
that the ongoing decline in manufacturing employment and the disappearance
of other routine jobs is causing the current low rates of employment.3 In ad1
We refer to computerisation as job automation by means of computer-controlled equipment.
2
This view finds support in a recent survey by the McKinsey Global Institute (MGI), showing
that 44 percent of firms which reduced their headcount since the financial crisis of 2008 had
done so by means of automation (MGI, 2011).
3
Because the core job tasks of manufacturing occupations follow well-defined repetitive
procedures, they can easily be codified in computer software and thus performed by computers
(Acemoglu and Autor, 2011).
2
dition to the computerisation of routine manufacturing tasks, Autor and Dorn
(2013) document a structural shift in the labour market, with workers reallocating their labour supply from middle-income manufacturing to low-income
service occupations. Arguably, this is because the manual tasks of service occupations are less susceptible to computerisation, as they require a higher degree
of flexibility and physical adaptability (Autor, et al., 2003; Goos and Manning,
2007; Autor and Dorn, 2013).
At the same time, with falling prices of computing, problem-solving skills
are becoming relatively productive, explaining the substantial employment growth
in occupations involving cognitive tasks where skilled labour has a comparative
advantage, as well as the persistent increase in returns to education (Katz and
Murphy, 1992; Acemoglu, 2002; Autor and Dorn, 2013). The title “Lousy and
Lovely Jobs”, of recent work by Goos and Manning (2007), thus captures the
essence of the current trend towards labour market polarization, with growing
employment in high-income cognitive jobs and low-income manual occupations, accompanied by a hollowing-out of middle-income routine jobs.
According to Brynjolfsson and McAfee (2011), the pace of technological innovation is still increasing, with more sophisticated software technologies disrupting labour markets by making workers redundant. What is striking
about the examples in their book is that computerisation is no longer confined
to routine manufacturing tasks. The autonomous driverless cars, developed by
Google, provide one example of how manual tasks in transport and logistics
may soon be automated. In the section “In Domain After Domain, Computers Race Ahead”, they emphasise how fast moving these developments have
been. Less than ten years ago, in the chapter “Why People Still Matter”, Levy
and Murnane (2004) pointed at the difficulties of replicating human perception,
asserting that driving in traffic is insusceptible to automation: “But executing a left turn against oncoming traffic involves so many factors that it is hard
to imagine discovering the set of rules that can replicate a driver’s behaviour
[. . . ]”. Six years later, in October 2010, Google announced that it had modified several Toyota Priuses to be fully autonomous (Brynjolfsson and McAfee,
2011).
To our knowledge, no study has yet quantified what recent technological
progress is likely to mean for the future of employment. The present study
intends to bridge this gap in the literature. Although there are indeed existing
3
useful frameworks for examining the impact of computers on the occupational
employment composition, they seem inadequate in explaining the impact of
technological trends going beyond the computerisation of routine tasks. Seminal work by Autor, et al. (2003), for example, distinguishes between cognitive
and manual tasks on the one hand, and routine and non-routine tasks on the
other. While the computer substitution for both cognitive and manual routine
tasks is evident, non-routine tasks involve everything from legal writing, truck
driving and medical diagnoses, to persuading and selling. In the present study,
we will argue that legal writing and truck driving will soon be automated, while
persuading, for instance, will not. Drawing upon recent developments in Engineering Sciences, and in particular advances in the fields of ML, including
Data Mining, Machine Vision, Computational Statistics and other sub-fields of
Artificial Intelligence, as well as MR, we derive additional dimensions required
to understand the susceptibility of jobs to computerisation. Needless to say,
a number of factors are driving decisions to automate and we cannot capture
these in full. Rather we aim, from a technological capabilities point of view,
to determine which problems engineers need to solve for specific occupations
to be automated. By highlighting these problems, their difficulty and to which
occupations they relate, we categorise jobs according to their susceptibility to
computerisation. The characteristics of these problems were matched to different occupational characteristics, using O∗NET data, allowing us to examine
the future direction of technological change in terms of its impact on the occupational composition of the labour market, but also the number of jobs at risk
should these technologies materialise.
The present study relates to two literatures. First, our analysis builds on the
labour economics literature on the task content of employment (Autor, et al.,
2003; Goos and Manning, 2007; Autor and Dorn, 2013). Based on defined
premises about what computers do, this literature examines the historical impact of computerisation on the occupational composition of the labour market. However, the scope of what computers do has recently expanded, and will
inevitably continue to do so (Brynjolfsson and McAfee, 2011; MGI, 2013).
Drawing upon recent progress in ML, we expand the premises about the tasks
computers are and will be suited to accomplish. Doing so, we build on the task
content literature in a forward-looking manner. Furthermore, whereas this literature has largely focused on task measures from the Dictionary of Occupational
4
Titles (DOT), last revised in 1991, we rely on the 2010 version of the DOT successor O∗NET – an online service developed for the US Department of Labor.4
Accordingly, O∗NET has the advantage of providing more recent information
on occupational work activities.
Second, our study relates to the literature examining the offshoring of information-based tasks to foreign worksites (Jensen and Kletzer, 2005; Blinder,
2009; Jensen and Kletzer, 2010; Oldenski, 2012; Blinder and Krueger, 2013).
This literature consists of different methodologies to rank and categorise occupations according to their susceptibility to offshoring. For example, using
O ∗NET data on the nature of work done in different occupations, Blinder (2009)
estimates that 22 to 29 percent of US jobs are or will be offshorable in the next
decade or two. These estimates are based on two defining characteristics of jobs
that cannot be offshored: (a) the job must be performed at a specific work loca-
tion; and (b) the job requires face-to-face personal communication. Naturally,
the characteristics of occupations that can be offshored are different from the
characteristics of occupations that can be automated. For example, the work of
cashiers, which has largely been substituted by self- service technology, must
be performed at specific work location and requires face-to-face contact. The
extent of computerisation is therefore likely to go beyond that of offshoring.
Hence, while the implementation of our methodology is similar to that of Blinder (2009), we rely on different occupational characteristics.
The remainder of this paper is structured as follows. In Section II, we review
the literature on the historical relationship between technological progress and
employment. Section III describes recent and expected future technological
developments. In Section IV, we describe our methodology, and in Section V,
we examine the expected impact of these technological developments on labour
market outcomes. Finally, in Section VI, we derive some conclusions.
II.
A HISTORY
OF TECHNOLOGICAL REVOLUTIONS AND EMPLOYMENT
The concern over technological unemployment is hardly a recent phenomenon.
Throughout history, the process of creative destruction, following technological inventions, has created enormous wealth, but also undesired disruptions.
As stressed by Schumpeter (1962), it was not the lack of inventive ideas that
4
An exception is Goos, et al. (2009).
5
set the boundaries for economic development, but rather powerful social and
economic interests promoting the technological status quo. This is nicely illustrated by the example of William Lee, inventing the stocking frame knitting
machine in 1589, hoping that it would relieve workers of hand-knitting. Seeking patent protection for his invention, he travelled to London where he had
rented a building for his machine to be viewed by Queen Elizabeth I. To his
disappointment, the Queen was more concerned with the employment impact
of his invention and refused to grant him a patent, claiming that: “Thou aimest
high, Master Lee. Consider thou what the invention could do to my poor subjects. It would assuredly bring to them ruin by depriving them of employment,
thus making them beggars” (cited in Acemoglu and Robinson, 2012, p. 182f).
Most likely the Queen’s concern was a manifestation of the hosiers’ guilds fear
that the invention would make the skills of its artisan members obsolete.5 The
guilds’ opposition was indeed so intense that William Lee had to leave Britain.
That guilds systematically tried to weaken market forces as aggregators to
maintain the technological status quo is persuasively argued by Kellenbenz
(1974, p. 243), stating that “guilds defended the interests of their members
against outsiders, and these included the inventors who, with their new equipment and techniques, threatened to disturb their members’ economic status.”6
As pointed out by Mokyr (1998, p. 11): “Unless all individuals accept the
“verdict” of the market outcome, the decision whether to adopt an innovation
is likely to be resisted by losers through non-market mechanism and political
activism.” Workers can thus be expected to resist new technologies, insofar that
they make their skills obsolete and irreversibly reduce their expected earnings.
The balance between job conservation and technological progress therefore, to
a large extent, reflects the balance of power in society, and how gains from
technological progress are being distributed.
The British Industrial Revolution illustrates this point vividly. While still
widely present on the Continent, the craft guild in Britain had, by the time of
5
The term artisan refers to a craftsman who engages in the entire production process of a
good, containing almost no division of labour. By guild we mean an association of artisans that
control the practice of their craft in a particular town.
6
There is an ongoing debate about the technological role of the guilds. Epstein (1998), for
example, has argued that they fulfilled an important role in the intergenerational transmission of
knowledge. Yet there is no immediate contradiction between such a role and their conservative
stand on technological progress: there are clear examples of guilds restraining the diffusion of
inventions (see, for example, Ogilvie, 2004).
6
the Glorious Revolution of 1688, declined and lost most of its political clout
(Nef, 1957, pp. 26 and 32). With Parliamentary supremacy established over
the Crown, legislation was passed in 1769 making the destruction of machinery
punishable by death (Mokyr, 1990, p. 257). To be sure, there was still resistance
to mechanisation. The “Luddite” riots between 1811 and 1816 were partly a
manifestation of the fear of technological change among workers as Parliament
revoked a 1551 law prohibiting the use of gig mills in the wool-finishing trade.
The British government however took an increasingly stern view on groups
attempting to halt technological progress and deployed 12,000 men against the
rioters (Mantoux, 2006, p. 403-8). The sentiment of the government towards
the destruction of machinery was explained by a resolution passed after the
Lancashire riots of 1779, stating that: “The sole cause of great riots was the
new machines employed in cotton manufacture; the country notwithstanding
has greatly benefited from their erection [and] destroying them in this country
would only be the means of transferring them to another [. . . ] to the detriment
of the trade of Britain” (cited in Mantoux, 2006, p. 403).
There are at least two possible explanations for the shift in attitudes towards
technological progress. First, after Parliamentary supremacy was established
over the Crown, the property owning classes became politically dominant in
Britain (North and Weingast, 1989). Because the diffusion of various manufacturing technologies did not impose a risk to the value of their assets, and some
property owners stood to benefit from the export of manufactured goods, the
artisans simply did not have the political power to repress them. Second, inventors, consumers and unskilled factory workers largely benefited from mechanisation (Mokyr, 1990, p. 256 and 258). It has even been argued that, despite
the employment concerns over mechanisation, unskilled workers have been the
greatest beneficiaries of the Industrial Revolution (Clark, 2008).7 While there
7
Various estimations of the living standards of workers in Britain during the industrialisation
exist in the literature. For example, Clark (2008) finds that real wages over the period 1760 to
1860 rose faster than GDP per capita. Further evidence provided by Lindert and Williamson
(1983) even suggests that real wages nearly doubled between 1820 and 1850. Feinstein (1998),
on the other hand, finds a much more moderate increase, with average working-class living
standards improving by less than 15 percent between 1770 and 1870. Finally, Allen (2009a)
finds that over the first half of the nineteenth century, the real wage stagnated while output per
worker expanded. After the mid nineteenth century, however, real wages began to grow in line
with productivity. While this implies that capital owners were the greatest beneficiaries of the
Industrial Revolution, there is at the same time consensus that average living standards largely
improved.
7
is contradictory evidence suggesting that capital owners initially accumulated
a growing share of national income (Allen, 2009a), there is equally evidence
of growing real wages (Lindert and Williamson, 1983; Feinstein, 1998). This
implies that although manufacturing technologies made the skills of artisans
obsolete, gains from technological progress were distributed in a manner that
gradually benefited a growing share of the labour force.8
An important feature of nineteenth century manufacturing technologies is
that they were largely “deskilling” – i.e. they substituted for skills through the
simplification of tasks (Braverman, 1974; Hounshell, 1985; James and Skinner,
1985; Goldin and Katz, 1998). The deskilling process occurred as the factory
system began to displace the artisan shop, and it picked up pace as production increasingly mechanized with the adoption of steam power (Goldin and
Sokoloff, 1982; Atack, et al., 2008a). Work that had previously been performed
by artisans was now decomposed into smaller, highly specialised, sequences,
requiring less skill, but more workers, to perform.9 Some innovations were
even designed to be deskilling. For example, Eli Whitney, a pioneer of interchangeable parts, described the objective of this technology as “to substitute
correct and effective operations of machinery for the skill of the artist which is
acquired only by long practice and experience; a species of skill which is not
possessed in this country to any considerable extent” (Habakkuk, 1962, p. 22).
Together with developments in continuous-flow production, enabling workers to be stationary while different tasks were moved to them, it was identical interchangeable parts that allowed complex products to be assembled from mass
produced individual components by using highly specialised machine tools to
8
The term skill is associated with higher levels of education, ability, or job training. Following Goldin and Katz (1998), we refer to technology-skill or capital-skill complementarity when
a new technology or physical capital complements skilled labour relative to unskilled workers.
9
The production of plows nicely illustrates the differences between the artisan shop and the
factory. In one artisan shop, two men spent 118 man-hours using hammers, anvils, chisels,
hatchets, axes, mallets, shaves and augers in 11 distinct operations to produce a plow. By
contrast, a mechanized plow factory employed 52 workers performing 97 distinct tasks, of
which 72 were assisted by steam power, to produce a plow in just 3.75 man-hours. The degree
of specialization was even greater in the production of men’s white muslin shirts. In the artisan
shop, one worker spent 1439 hours performing 25 different tasks to produce 144 shirts. In the
factory, it took 188 man-hours to produce the same quantity, engaging 230 different workers
performing 39 different tasks, of which more than half required steam power. The workers
involved included cutters, turners and trimmers, as well as foremen and forewomen, inspectors,
errand boys, an engineer, a fireman, and a watchman (US Department of Labor, 1899).
8
a sequence of operations.10 Yet while the first assembly-line was documented
in 1804, it was not until the late nineteenth century that continuous-flow processes started to be adopted on a larger scale, which enabled corporations such
as the Ford Motor Company to manufacture the T-Ford at a sufficiently low
price for it to become the people’s vehicle (Mokyr, 1990, p. 137). Crucially,
the new assembly line introduced by Ford in 1913 was specifically designed for
machinery to be operated by unskilled workers (Hounshell, 1985, p. 239). Furthermore, what had previously been a one-man job was turned into a 29-man
worker operation, reducing the overall work time by 34 percent (Bright, 1958).
The example of the Ford Motor Company thus underlines the general pattern
observed in the nineteenth century, with physical capital providing a relative
complement to unskilled labour, while substituting for relatively skilled artisans (James and Skinner, 1985; Louis and Paterson, 1986; Brown and Philips,
1986; Atack, et al., 2004).11 Hence, as pointed out by Acemoglu (2002, p. 7):
“the idea that technological advances favor more skilled workers is a twentieth
century phenomenon.” The conventional wisdom among economic historians,
in other words, suggests a discontinuity between the nineteenth and twentieth
century in the impact of capital deepening on the relative demand for skilled
labour.
The modern pattern of capital-skill complementarity gradually emerged in
the late nineteenth century, as manufacturing production shifted to increasingly
mechanised assembly lines. This shift can be traced to the switch to electricity
from steam and water-power which, in combination with continuous-process
10
These machines were sequentially implemented until the production process was completed. Over time, such machines became much cheaper relative to skilled labor. As a result,
production became much more capital intensive (Hounshell, 1985).
11
Williamson and Lindert (1980), on the other hand, find a relative rise in wage premium of
skilled labour over the period 1820 to 1860, which they partly attribute to capital deepening.
Their claim of growing wage inequality over this period has, however, been challenged (Margo,
2000). Yet seen over the long-run, a more refined explanation is that the manufacturing share
of the labour force in the nineteenth century hollowed out. This is suggested by recent findings,
revealing a decline of middle-skill artisan jobs in favour of both high-skill white collar workers
and low-skill operatives (Gray, 2013; Katz and Margo, 2013). Furthermore, even if the share
of operatives was increasing due to organizational change within manufacturing and overall
manufacturing growth, it does not follow that the share of unskilled labor was rising in the
aggregate economy, because some of the growth in the share of operatives may have come
at the expense of a decrease in the share of workers employed as low-skilled farm workers in
agriculture (Katz and Margo, 2013). Nevertheless, this evidence is consistent with the literature
showing that relatively skilled artisans were replaced by unskilled factory workers, suggesting
that technological change in manufacturing was deskilling.
9
and batch production methods, reduced the demand for unskilled manual workers in many hauling, conveying, and assembly tasks, but increased the demand
for skills (Goldin and Katz, 1998). In short, while factory assembly lines, with
their extreme division of labour, had required vast quantities of human operatives, electrification allowed many stages of the production process to be automated, which in turn increased the demand for relatively skilled blue-collar
production workers to operate the machinery. In addition, electrification contributed to a growing share of white-collar nonproduction workers (Goldin and
Katz, 1998). Over the course of the nineteenth century, establishments became
larger in size as steam and water power technologies improved, allowing them
to adopt powered machinery to realize productivity gains through the combination of enhanced division of labour and higher capital intensity (Atack, et al.,
2008a). Furthermore, the transport revolution lowered costs of shipping goods
domestically and internationally as infrastructure spread and improved (Atack,
et al., 2008b). The market for artisan goods early on had largely been confined
to the immediate surrounding area because transport costs were high relative to
the value of the goods produced. With the transport revolution, however, market
size expanded, thereby eroding local monopoly power, which in turn increased
competition and compelled firms to raise productivity through mechanisation.
As establishments became larger and served geographically expended markets,
managerial tasks increased in number and complexity, requiring more managerial and clerking employees (Chandler, 1977). This pattern was, by the turn of
the twentieth century, reinforced by electrification, which not only contributed
to a growing share of relatively skilled blue-collar labour, but also increased the
demand for white-collar workers (Goldin and Katz, 1998), who tended to have
higher educational attainment (Allen, 2001).12
Since electrification, the story of the twentieth century has been the race between education and technology (Goldin and Katz, 2009). The US high school
movement coincided with the first industrial revolution of the office (Goldin
and Katz, 1995). While the typewriter was invented in the 1860s, it was not introduced in the office until the early twentieth century, when it entered a wave
12
Most likely, the growing share of white-collar workers increased the element of human
interaction in employment. Notably, Michaels, et al. (2013) find that the increase in the employment share of interactive occupations, going hand in hand with an increase in their relative
wage bill share, was particularly strong between 1880 and 1930, which is a period of rapid
change in communication and transport technology.
10
of mechanisation, with dictaphones, calculators, mimeo machines, address machines, and the predecessor of the computer – the keypunch (Beniger, 1986;
Cortada, 2000). Importantly, these office machines reduced the cost of information processing tasks and increased the demand for the complementary factor –
i.e. educated office workers. Yet the increased supply of educated office workers, following the high school movement, was associated with a sharp decline
in the wage premium of clerking occupations relative to production workers
(Goldin and Katz, 1995). This was, however, not the result of deskilling technological change. Clerking workers were indeed relatively educated. Rather, it
was the result of the supply of educated workers outpacing the demand for their
skills, leading educational wage differentials to compress.
While educational wage differentials in the US narrowed from 1915 to 1980
(Goldin and Katz, 2009), both educational wage differentials and overall wage
inequality have increased sharply since the 1980s in a number of countries
(Krueger, 1993; Murphy, et al., 1998; Atkinson, 2008; Goldin and Katz, 2009).
Although there are clearly several variables at work, consensus is broad that
this can be ascribed to an acceleration in capital-skill complementarity, driven
by the adoption of computers and information technology (Krueger, 1993; Autor, et al., 1998; Bresnahan, et al., 2002). What is commonly referred to as the
Computer Revolution began with the first commercial uses of computers around
1960 and continued through the development of the Internet and e-commerce
in the 1990s. As the cost per computation declined at an annual average of 37
percent between 1945 and 1980 (Nordhaus, 2007), telephone operators were
made redundant, the first industrial robot was introduced by General Motors
in the 1960s, and in the 1970s airline reservations systems led the way in selfservice technology (Gordon, 2012). During the 1980s and 1990s, computing
costs declined even more rapidly, on average by 64 percent per year, accompanied by a surge in computational power (Nordhaus, 2007).13 At the same time,
bar-code scanners and cash machines were spreading across the retail and financial industries, and the first personal computers were introduced in the early
1980s, with their word processing and spreadsheet functions eliminating copy
typist occupations and allowing repetitive calculations to be automated (Gordon, 2012). This substitution for labour marks a further important reversal.
13
Computer power even increased 18 percent faster on annual basis than predicted by
Moore’s Law, implying a doubling every two years (Nordhaus, 2007).
11
The early twentieth century office machines increased the demand for clerking
workers (Chandler, 1977; Goldin and Katz, 1995). In a similar manner, computerisation augments demand for such tasks, but it also permits them to be
automated (Autor, et al., 2003).
The Computer Revolution can go some way in explaining the growing wage
inequality of the past decades. For example, Krueger (1993) finds that workers using a computer earn roughly earn 10 to 15 percent more than others, but
also that computer use accounts for a substantial share of the increase in the
rate of return to education. In addition, more recent studies find that computers
have caused a shift in the occupational structure of the labour market. Autor
and Dorn (2013), for example, show that as computerisation erodes wages for
labour performing routine tasks, workers will reallocate their labour supply to
relatively low-skill service occupations. More specifically, between 1980 and
2005, the share of US labour hours in service occupations grew by 30 percent
after having been flat or declining in the three prior decades. Furthermore, net
changes in US employment were U-shaped in skill level, meaning that the lowest and highest job-skill quartile expanded sharply with relative employment
declines in the middle of the distribution.
The expansion in high-skill employment can be explained by the falling
price of carrying out routine tasks by means of computers, which complements
more abstract and creative services. Seen from a production function perspective, an outward shift in the supply of routine informational inputs increases the
marginal productivity of workers they are demanded by. For example, text and
data mining has improved the quality of legal research as constant access to
market information has improved the efficiency of managerial decision-making
– i.e. tasks performed by skilled workers at the higher end of the income distribution. The result has been an increasingly polarised labour market, with
growing employment in high-income cognitive jobs and low-income manual
occupations, accompanied by a hollowing-out of middle-income routine jobs.
This is a pattern that is not unique to the US and equally applies to a number of
developed economies (Goos, et al., 2009).14
14
While there is broad consensus that computers substituting for workers in routine-intensive
tasks has driven labour market polarisation over the past decades, there are, indeed, alternative
explanations. For example, technological advances in computing have dramatically lowered the
cost of leaving information-based tasks to foreign worksites (Jensen and Kletzer, 2005; Blinder,
2009; Jensen and Kletzer, 2010; Oldenski, 2012; Blinder and Krueger, 2013). The decline in
12
How technological progress in the twenty-first century will impact on labour
market outcomes remains to be seen. Throughout history, technological progress
has vastly shifted the composition of employment, from agriculture and the
artisan shop, to manufacturing and clerking, to service and management occupations. Yet the concern over technological unemployment has proven to
be exaggerated. The obvious reason why this concern has not materialised
relates to Ricardo’s famous chapter on machinery, which suggests that laboursaving technology reduces the demand for undifferentiated labour, thus leading
to technological unemployment (Ricardo, 1819). As economists have long understood, however, an invention that replaces workers by machines will have
effects on all product and factor markets. An increase in the efficiency of production which reduces the price of one good, will increase real income and
thus increase demand for other goods. Hence, in short, technological progress
has two competing effects on employment (Aghion and Howitt, 1994). First, as
technology substitutes for labour, there is a destruction effect, requiring workers
to reallocate their labour supply; and second, there is the capitalisation effect, as
more companies enter industries where productivity is relatively high, leading
employment in those industries to expand.
Although the capitalisation effect has been predominant historically, our
discovery of means of economising the use of labour can outrun the pace at
which we can find new uses for labour, as Keynes (1933) pointed out. The reason why human labour has prevailed relates to its ability to adopt and acquire
new skills by means of education (Goldin and Katz, 2009). Yet as computerisation enters more cognitive domains this will become increasingly challenging
(Brynjolfsson and McAfee, 2011). Recent empirical findings are therefore particularly concerning. For example, Beaudry, et al. (2013) document a decline
in the demand for skill over the past decade, even as the supply of workers with
higher education has continued to grow. They show that high-skilled workers have moved down the occupational ladder, taking on jobs traditionally performed by low-skilled workers, pushing low-skilled workers even further down
the occupational ladder and, to some extent, even out of the labour force. This
the routine-intensity of employment is thus likely to result from a combination of offshoring
and automation. Furthermore, there is evidence suggesting that improvements in transport and
communication technology have augmented occupations involving human interaction, spanning across both cognitive and manual tasks (Michaels, et al., 2013). These explanations are
nevertheless equally related to advance in computing and communications technology.
13
raises questions about: (a) the ability of human labour to win the race against
technology by means of education; and (b) the potential extent of technological unemployment, as an increasing pace of technological progress will cause
higher job turnover, resulting in a higher natural rate of unemployment (Lucas
and Prescott, 1974; Davis and Haltiwanger, 1992; Pissarides, 2000). While the
present study is limited to examining the destruction effect of technology, it
nevertheless provides a useful indication of the job growth required to counterbalance the jobs at risk over the next decades.
III.
T HE
TECHNOLOGICAL REVOLUTIONS OF THE TWENTY- FIRST CENTURY
The secular price decline in the real cost of computing has created vast economic incentives for employers to substitute labour for computer capital.15 Yet
the tasks computers are able to perform ultimately depend upon the ability of
a programmer to write a set of procedures or rules that appropriately direct the
technology in each possible contingency. Computers will therefore be relatively
productive to human labour when a problem can be specified – in the sense that
the criteria for success are quantifiable and can readily be evaluated (Acemoglu
and Autor, 2011). The extent of job computerisation will thus be determined
by technological advances that allow engineering problems to be sufficiently
specified, which sets the boundaries for the scope of computerisation. In this
section, we examine the extent of tasks computer-controlled equipment can be
expected to perform over the next decades. Doing so, we focus on advances
in fields related to Machine Learning (ML), including Data Mining, Machine
Vision, Computational Statistics and other sub-fields of Artificial Intelligence
(AI), in which efforts are explicitly dedicated to the development of algorithms
that allow cognitive tasks to be automated. In addition, we examine the application of ML technologies in Mobile Robotics (MR), and thus the extent of
computerisation in manual tasks.
Our analysis builds on the task categorisation of Autor, et al. (2003), which
distinguishes between workplace tasks using a two-by-two matrix, with routine
versus non-routine tasks on one axis, and manual versus cognitive tasks on the
other. In short, routine tasks are defined as tasks that follow explicit rules that
15
We refer to computer capital as accumulated computers and computer-controlled equipment by means of capital deepening.
14
can be accomplished by machines, while non-routine tasks are not sufficiently
well understood to be specified in computer code. Each of these task categories can, in turn, be of either manual or cognitive nature – i.e. they relate to
physical labour or knowledge work. Historically, computerisation has largely
been confined to manual and cognitive routine tasks involving explicit rulebased activities (Autor and Dorn, 2013; Goos, et al., 2009). Following recent
technological advances, however, computerisation is now spreading to domains
commonly defined as non-routine. The rapid pace at which tasks that were defined as non-routine only a decade ago have now become computerisable is
illustrated by Autor, et al. (2003), asserting that: “Navigating a car through city
traffic or deciphering the scrawled handwriting on a personal check – minor
undertakings for most adults – are not routine tasks by our definition.” Today,
the problems of navigating a car and deciphering handwriting are sufficiently
well understood that many related tasks can be specified in computer code and
automated (Veres, et al., 2011; Plötz and Fink, 2009).
Recent technological breakthroughs are, in large part, due to efforts to turn
non-routine tasks into well-defined problems. Defining such problems is helped
by the provision of relevant data: this is highlighted in the case of handwriting
recognition by Plötz and Fink (2009). The success of an algorithm for handwriting recognition is difficult to quantify without data to test on – in particular,
determining whether an algorithm performs well for different styles of writing requires data containing a variety of such styles. That is, data is required
to specify the many contingencies a technology must manage in order to form
an adequate substitute for human labour. With data, objective and quantifiable
measures of the success of an algorithm can be produced, which aid the continual improvement of its performance relative to humans.
As such, technological progress has been aided by the recent production
of increasingly large and complex datasets, known as big data.16 For instance,
with a growing corpus of human-translated digitalised text, the success of a
machine translator can now be judged by its accuracy in reproducing observed
translations. Data from United Nations documents, which are translated by hu16
Predictions by Cisco Systems suggest that the Internet traffic in 2016 will be around 1
zettabyte (1 × 1021 bytes) (Cisco, 2012). In comparison, the information contained in all books
worldwide is about 480 terabytes (5 × 1014 bytes), and a text transcript of all the words ever
spoken by humans would represent about 5 exabytes (5 × 1018 bytes) (UC Berkeley School of
Information, 2003).
15
man experts into six languages, allow Google Translate to monitor and improve
the performance of different machine translation algorithms (Tanner, 2007).
Further, ML algorithms can discover unexpected similarities between old
and new data, aiding the computerisation of tasks for which big data has newly
become available. As a result, computerisation is no longer confined to routine tasks that can be written as rule-based software queries, but is spreading
to every non-routine task where big data becomes available (Brynjolfsson and
McAfee, 2011). In this section, we examine the extent of future computerisation beyond routine tasks.
III.A. Computerisation in non-routine cognitive tasks
With the availability of big data, a wide range of non-routine cognitive tasks
are becoming computerisable. That is, further to the general improvement in
technological progress due to big data, algorithms for big data are rapidly entering domains reliant upon storing or accessing information. The use of big data
is afforded by one of the chief comparative advantages of computers relative
to human labor: scalability. Little evidence is required to demonstrate that, in
performing the task of laborious computation, networks of machines scale better than human labour (Campbell-Kelly, 2009). As such, computers can better
manage the large calculations required in using large datasets. ML algorithms
running on computers are now, in many cases, better able to detect patterns in
big data than humans.
Computerisation of cognitive tasks is also aided by another core comparative advantage of algorithms: their absence of some human biases. An algorithm can be designed to ruthlessly satisfy the small range of tasks it is given.
Humans, in contrast, must fulfill a range of tasks unrelated to their occupation,
such as sleeping, necessitating occasional sacrifices in their occupational performance (Kahneman, et al., 1982). The additional constraints under which
humans must operate manifest themselves as biases. Consider an example of
human bias: Danziger, et al. (2011) demonstrate that experienced Israeli judges
are substantially more generous in their rulings following a lunch break. It can
thus be argued that many roles involving decision-making will benefit from
impartial algorithmic solutions.
Fraud detection is a task that requires both impartial decision making and
the ability to detect trends in big data. As such, this task is now almost com16
pletely automated (Phua, et al., 2010). In a similar manner, the comparative
advantages of computers are likely to change the nature of work across a wide
range of industries and occupations.
In health care, diagnostics tasks are already being computerised. Oncologists at Memorial Sloan-Kettering Cancer Center are, for example, using IBM’s
Watson computer to provide chronic care and cancer treatment diagnostics.
Knowledge from 600,000 medical evidence reports, 1.5 million patient records
and clinical trials, and two million pages of text from medical journals, are used
for benchmarking and pattern recognition purposes. This allows the computer
to compare each patient’s individual symptoms, genetics, family and medication history, etc., to diagnose and develop a treatment plan with the highest
probability of success (Cohn, 2013).
In addition, computerisation is entering the domains of legal and financial
services. Sophisticated algorithms are gradually taking on a number of tasks
performed by paralegals, contract and patent lawyers (Markoff, 2011). More
specifically, law firms now rely on computers that can scan thousands of legal
briefs and precedents to assist in pre-trial research. A frequently cited example is Symantec’s Clearwell system, which uses language analysis to identify
general concepts in documents, can present the results graphically, and proved
capable of analysing and sorting more than 570,000 documents in two days
(Markoff, 2011).
Furthermore, the improvement of sensing technology has made sensor data
one of the most prominent sources of big data (Ackerman and Guizzo, 2011).
Sensor data is often coupled with new ML fault- and anomaly-detection algorithms to render many tasks computerisable. A broad class of examples can be
found in condition monitoring and novelty detection, with technology substituting for closed-circuit TV (CCTV) operators, workers examining equipment
defects, and clinical staff responsible for monitoring the state of patients in intensive care. Here, the fact that computers lack human biases is of great value:
algorithms are free of irrational bias, and their vigilance need not be interrupted
by rest breaks or lapses of concentration. Following the declining costs of digital sensing and actuation, ML approaches have successfully addressed condition
monitoring applications ranging from batteries (Saha, et al., 2007), to aircraft
engines (King, et al., 2009), water quality (Osborne, et al., 2012) and intensive
care units (ICUs) (Clifford and Clifton, 2012; Clifton, et al., 2012). Sensors can
17
equally be placed on trucks and pallets to improve companies’ supply chain
management, and used to measure the moisture in a field of crops to track the
flow of water through utility pipes. This allows for automatic meter reading,
eliminating the need for personnel to gather such information. For example,
the cities of Doha, São Paulo, and Beijing use sensors on pipes, pumps, and
other water infrastructure to monitor conditions and manage water loss, reducing leaks by 40 to 50 percent. In the near future, it will be possible to place inexpensive sensors on light poles, sidewalks, and other public property to capture
sound and images, likely reducing the number of workers in law enforcement
(MGI, 2013).
Advances in user interfaces also enable computers to respond directly to
a wider range of human requests, thus augmenting the work of highly skilled
labour, while allowing some types of jobs to become fully automated. For example, Apple’s Siri and Google Now rely on natural user interfaces to recognise
spoken words, interpret their meanings, and act on them accordingly. Moreover, a company called SmartAction now provides call computerisation solutions that use ML technology and advanced speech recognition to improve upon
conventional interactive voice response systems, realising cost savings of 60 to
80 percent over an outsourced call center consisting of human labour (CAA,
2012). Even education, one of the most labour intensive sectors, will most
likely be significantly impacted by improved user interfaces and algorithms
building upon big data. The recent growth in MOOCs (Massive Open Online
Courses) has begun to generate large datasets detailing how students interact
on forums, their diligence in completing assignments and viewing lectures, and
their ultimate grades (Simonite, 2013; Breslow, et al., 2013). Such information,
together with improved user interfaces, will allow for ML algorithms that serve
as interactive tutors, with teaching and assessment strategies statistically calibrated to match individual student needs (Woolf, 2010). Big data analysis will
also allow for more effective predictions of student performance, and for their
suitability for post-graduation occupations. These technologies can equally be
implemented in recruitment, most likely resulting in the streamlining of human
resource (HR) departments.
Occupations that require subtle judgement are also increasingly susceptible
to computerisation. To many such tasks, the unbiased decision making of an algorithm represents a comparative advantage over human operators. In the most
18
challenging or critical applications, as in ICUs, algorithmic recommendations
may serve as inputs to human operators; in other circumstances, algorithms
will themselves be responsible for appropriate decision-making. In the financial sector, such automated decision-making has played a role for quite some
time. AI algorithms are able to process a greater number of financial announcements, press releases, and other information than any human trader, and then
act faster upon them (Mims, 2010). Services like Future Advisor similarly use
AI to offer personalised financial advice at larger scale and lower cost. Even
the work of software engineers may soon largely be computerisable. For example, advances in ML allow a programmer to leave complex parameter and
design choices to be appropriately optimised by an algorithm (Hoos, 2012). Algorithms can further automatically detect bugs in software (Hangal and Lam,
2002; Livshits and Zimmermann, 2005; Kim, et al., 2008), with a reliability
that humans are unlikely to match. Big databases of code also offer the eventual
prospect of algorithms that learn how to write programs to satisfy specifications
provided by a human. Such an approach is likely to eventually improve upon
human programmers, in the same way that human-written compilers eventually
proved inferior to automatically optimised compilers. An algorithm can better keep the whole of a program in working memory, and is not constrained to
human-intelligible code, allowing for holistic solutions that might never occur
to a human. Such algorithmic improvements over human judgement are likely
to become increasingly common.
Although the extent of these developments remains to be seen, estimates by
MGI (2013) suggests that sophisticated algorithms could substitute for approximately 140 million full-time knowledge workers worldwide. Hence, while
technological progress throughout economic history has largely been confined
to the mechanisation of manual tasks, requiring physical labour, technological
progress in the twenty-first century can be expected to contribute to a wide
range of cognitive tasks, which, until now, have largely remained a human
domain. Of course, many occupations being affected by these developments
are still far from fully computerisable, meaning that the computerisation of
some tasks will simply free-up time for human labour to perform other tasks.
Nonetheless, the trend is clear: computers increasingly challenge human labour
in a wide range of cognitive tasks (Brynjolfsson and McAfee, 2011).
19
III.B. Computerisation in non-routine manual tasks
Mobile robotics provides a means of directly leveraging
ML
technologies to
aid the computerisation of a growing scope of manual tasks. The continued
technological development of robotic hardware is having notable impact upon
employment: over the past decades, industrial robots have taken on the routine tasks of most operatives in manufacturing. Now, however, more advanced
robots are gaining enhanced sensors and manipulators, allowing them to perform non-routine manual tasks. For example, General Electric has recently developed robots to climb and maintain wind turbines, and more flexible surgical
robots with a greater range of motion will soon perform more types of operations (Robotics-VO, 2013). In a similar manner, the computerisation of logistics is being aided by the increasing cost-effectiveness of highly instrumented
and computerised cars. Mass-production vehicles, such as the Nissan LEAF,
contain on-board computers and advanced telecommunication equipment that
render the car a potentially fly-by-wire robot.17 Advances in sensor technology mean that vehicles are likely to soon be augmented with even more advanced suites of sensors. These will permit an algorithmic vehicle controller to
monitor its environment to a degree that exceeds the capabilities of any human
driver: they have the ability to simultaneously look both forwards and backwards, can natively integrate camera, GPS and LIDAR data, and are not subject
to distraction. Algorithms are thus potentially safer and more effective drivers
than humans.
The big data provided by these improved sensors are offering solutions to
many of the engineering problems that had hindered robotic development in
the past. In particular, the creation of detailed three dimensional maps of road
networks has enabled autonomous vehicle navigation; most notably illustrated
by Google’s use of large, specialised datasets collected by its driverless cars
(Guizzo, 2011). It is now completely feasible to store representations of the
entire road network on-board a car, dramatically simplifying the navigation
problem. Algorithms that could perform navigation throughout the changing
seasons, particularly after snowfall, have been viewed as a substantial challenge. However, the big data approach can answer this by storing records from
the last time snow fell, against which the vehicle’s current environment can
17
A fly-by-wire robot is a robot that is controllable by a remote computer.
20
be compared (Churchill and Newman, 2012). ML approaches have also been
developed to identify unprecedented changes to a particular piece of the road
network, such as roadworks (Mathibela, et al., 2012). This emerging technology will affect a variety of logistics jobs. Agricultural vehicles, forklifts
and cargo-handling vehicles are imminently automatable, and hospitals are already employing autonomous robots to transport food, prescriptions and samples (Bloss, 2011). The computerisation of mining vehicles is further being
pursued by companies such as Rio Tinto, seeking to replace labour in Australian mine-sites.18
With improved sensors, robots are capable of producing goods with higher
quality and reliability than human labour. For example, El Dulze, a Spanish
food processor, now uses robotics to pick up heads of lettuce from a conveyor belt, rejecting heads that do not comply with company standards. This
is achieved by measuring their density and replacing them on the belt (IFR,
2012a). Advanced sensors further allow robots to recognise patterns. Baxter, a
22,000 USD general-purpose robot, provides a well-known example. The robot
features an LCD display screen displaying a pair of eyes that take on different expressions depending on the situation. When the robot is first installed or
needs to learn a new pattern, no programming is required. A human worker
simply guides the robot arms through the motions that will be needed for the
task. Baxter then memorises these patterns and can communicate that it has understood its new instructions. While the physical flexibility of Baxter is limited
to performing simple operations such as picking up objects and moving them,
different standard attachments can be installed on its arms, allowing Baxter to
perform a relatively broad scope of manual tasks at low cost (MGI, 2013).
Technological advances are contributing to declining costs in robotics. Over
the past decades, robot prices have fallen about 10 percent annually and are
expected to decline at an even faster pace in the near future (MGI, 2013). Industrial robots, with features enabled by machine vision and high-precision
dexterity, which typically cost 100,000 to 150,000 USD, will be available for
50,000 to 75,000 USD in the next decade, with higher levels of intelligence
and additional capabilities (IFR, 2012b). Declining robot prices will inevitably
place them within reach of more users. For example, in China, employers are
18
Rio Tinto’s computerisation efforts are advertised at http://www.mineofthefuture.com.au.
21
increasingly incentivised to substitute robots for labour, as wages and living
standards are rising – Foxconn, a Chinese contract manufacturer that employs
1.2 million workers, is now investing in robots to assemble products such as
the Apple iPhone (Markoff, 2012). According to the International Federation
of Robotics, robot sales in China grew by more than 50 percent in 2011 and are
expected to increase further. Globally, industrial robot sales reached a record
166,000 units in 2011, a 40 percent year-on-year increase (IFR, 2012b). Most
likely, there will be even faster growth ahead as low-priced general-purpose
models, such as Baxter, are adopted in simple manufacturing and service work.
Expanding technological capabilities and declining costs will make entirely
new uses for robots possible. Robots will likely continue to take on an increasing set of manual tasks in manufacturing, packing, construction, maintenance,
and agriculture. In addition, robots are already performing many simple service tasks such as vacuuming, mopping, lawn mowing, and gutter cleaning –
the market for personal and household service robots is growing by about 20
percent annually (MGI, 2013). Meanwhile, commercial service robots are now
able to perform more complex tasks in food preparation, health care, commercial cleaning, and elderly care (Robotics-VO, 2013). As robot costs decline and
technological capabilities expand, robots can thus be expected to gradually substitute for labour in a wide range of low-wage service occupations, where most
job growth has occurred over the past decades (Autor and Dorn, 2013). This
means that many low-wage manual jobs that have been previously protected
from computerisation could diminish over time.
US
III.C.
The task model revisited
The task model of Autor, et al. (2003) has delivered intuitive and accurate
predictions in that: (a) computers are more substitutable for human labour in
routine relative to non-routine tasks; and (b) a greater intensity of routine inputs increases the marginal productivity of non-routine inputs. Accordingly,
computers have served as a substitute for labour for many routine tasks, while
exhibiting strong complementarities with labour performing cognitive non-routine tasks.19 Yet the premises about what computers do have recently expanded.
Computer capital can now equally substitute for a wide range of tasks com19
The model does not predict any substantial substitution or complementarity with nonroutine manual tasks.
22
monly defined as non-routine (Brynjolfsson and McAfee, 2011), meaning that
the task model will not hold in predicting the impact of computerisation on
the task content of employment in the twenty-first century. While focusing on
the substitution effects of recent technological progress, we build on the task
model by deriving several factors that we expect will determine the extent of
computerisation in non-routine tasks.
The task model assumes for tractability an aggregate, constant-returns-toscale, Cobb-Douglas production function of the form
(1)
Q = (LS + C)1−β LβNS ,
β ∈ [0, 1],
where LS and LNS are susceptible and non-susceptible labor inputs and C is
computer capital. Computer capital is supplied perfectly elastically at market
price per efficiency unit, where the market price is falling exogenously with
time due to technological progress. It further assumes income-maximizing
workers, with heterogeneous productivity endowments in both susceptible and
non-susceptible tasks. Their task supply will respond elastically to relative
wage levels, meaning that workers will reallocate their labour supply according
to their comparative advantage as in Roy (1951). With expanding computational capabilities, resulting from technological advances, and a falling market
price of computing, workers in susceptible tasks will thus reallocate to nonsusceptible tasks.
The above described simple model differs from the task model of Autor,
et al. (2003), in that LNS is not confined to routine labour inputs. This is because recent developments in ML and MR, building upon big data, allow for
pattern recognition, and thus enable computer capital to rapidly substitute for
labour across a wide range of non-routine tasks. Yet some inhibiting engineering bottlenecks to computerisation persist. Beyond these bottlenecks, however,
we argue that it is largely already technologically possible to automate almost
any task, provided that sufficient amounts of data are gathered for pattern recognition. Our model thus predicts that the pace at which these bottlenecks can be
overcome will determine the extent of computerisation in the twenty-first century.
Hence, in short, while the task model predicts that computers for labour
23
substitution will be confined to routine tasks, our model predicts that computerisation can be extended to any non-routine task that is not subject to any engineering bottlenecks to computerisation. These bottlenecks thus set the boundaries for the computerisation of non-routine tasks. Drawing upon the ML and
MR literature, and a workshop held at the Oxford University Engineering Sciences Department, we identify several engineering bottlenecks, corresponding
to three task categories. According to these findings, non-susceptible labor inputs can be described as,
(2)
LNS =
n
X
LPM,i + LC,i + LSI,i

i=1
where LPM , LC and LSI are labour inputs into perception and manipulation
tasks, creative intelligence tasks, and and social intelligence tasks.
We note that some related engineering bottlenecks can be partially alleviated by the simplification of tasks. One generic way of achieving this is to reduce the variation between task iterations. As a prototypical example, consider
the factory assembly line, turning the non-routine tasks of the artisan shop into
repetitive routine tasks performed by unskilled factory workers. A more recent
example is the computerisation of non-routine manual tasks in construction.
On-site construction tasks typically demand a high degree of adaptability, so
as to accommodate work environments that are typically irregularly laid out,
and vary according to weather. Prefabrication, in which the construction object
is partially assembled in a factory before being transported to the construction
site, provides a way of largely removing the requirement for adaptability. It allows many construction tasks to be performed by robots under controlled conditions that eliminate task variability – a method that is becoming increasingly
widespread, particularly in Japan (Barlow and Ozaki, 2005; Linner and Bock,
2012). The extent of computerisation in the twenty-first century will thus partly
depend on innovative approaches to task restructuring. In the remainder of this
section we examine the engineering bottlenecks related to the above mentioned
task categories, each in turn.
Perception and manipulation tasks. Robots are still unable to match the
depth and breadth of human perception. While basic geometric identification is
24
reasonably mature, enabled by the rapid development of sophisticated sensors
and lasers, significant challenges remain for more complex perception tasks,
such as identifying objects and their properties in a cluttered field of view. As
such, tasks that relate to an unstructured work environment can make jobs less
susceptible to computerisation. For example, most homes are unstructured, requiring the identification of a plurality of irregular objects and containing many
cluttered spaces which inhibit the mobility of wheeled objects. Conversely, supermarkets, factories, warehouses, airports and hospitals have been designed
for large wheeled objects, making it easier for robots to navigate in performing non-routine manual tasks. Perception problems can, however, sometimes
be sidestepped by clever task design. For example, Kiva Systems, acquired by
Amazon.com in 2012, solved the problem of warehouse navigation by simply
placing bar-code stickers on the floor, informing robots of their precise location
(Guizzo, 2008).
The difficulty of perception has ramifications for manipulation tasks, and,
in particular, the handling of irregular objects, for which robots are yet to reach
human levels of aptitude. This has been evidenced in the development of robots
that interact with human objects and environments. While advances have been
made, solutions tend to be unreliable over the myriad small variations on a single task, repeated thousands of times a day, that many applications require. A
related challenge is failure recovery – i.e. identifying and rectifying the mistakes of the robot when it has, for example, dropped an object. Manipulation is also limited by the difficulties of planning out the sequence of actions
required to move an object from one place to another. There are yet further
problems in designing manipulators that, like human limbs, are soft, have compliant dynamics and provide useful tactile feedback. Most industrial manipulation makes uses of workarounds to these challenges (Brown, et al., 2010),
but these approaches are nonetheless limited to a narrow range of tasks. The
main challenges to robotic computerisation, perception and manipulation, thus
largely remain and are unlikely to be fully resolved in the next decade or two
(Robotics-VO, 2013).
Creative intelligence tasks. The psychological processes underlying human
creativity are difficult to specify. According to Boden (2003), creativity is the
ability to come up with ideas or artifacts that are novel and valuable. Ideas, in a
25
broader sense, include concepts, poems, musical compositions, scientific theories, cooking recipes and jokes, whereas artifacts are objects such as paintings,
sculptures, machinery, and pottery. One process of creating ideas (and similarly for artifacts) involves making unfamiliar combinations of familiar ideas,
requiring a rich store of knowledge. The challenge here is to find some reliable
means of arriving at combinations that “make sense.” For a computer to make a
subtle joke, for example, would require a database with a richness of knowledge
comparable to that of humans, and methods of benchmarking the algorithm’s
subtlety.
In principle, such creativity is possible and some approaches to creativity
already exist in the literature. Duvenaud, et al. (2013) provide an example of
automating the core creative task required in order to perform statistics, that
of designing models for data. As to artistic creativity, AARON, a drawingprogram, has generated thousands of stylistically-similar line-drawings, which
have been exhibited in galleries worldwide. Furthermore, David Cope’s EMI
software composes music in many different styles, reminiscent of specific human composers.
In these and many other applications, generating novelty is not particularly
difficult. Instead, the principal obstacle to computerising creativity is stating
our creative values sufficiently clearly that they can be encoded in an program
(Boden, 2003). Moreover, human values change over time and vary across
cultures. Because creativity, by definition, involves not only novelty but value,
and because values are highly variable, it follows that many arguments about
creativity are rooted in disagreements about value. Thus, even if we could
identify and encode our creative values, to enable the computer to inform and
monitor its own activities accordingly, there would still be disagreement about
whether the computer appeared to be creative. In the absence of engineering
solutions to overcome this problem, it seems unlikely that occupations requiring
a high degree of creative intelligence will be automated in the next decades.
Social intelligence tasks. Human social intelligence is important in a wide
range of work tasks, such as those involving negotiation, persuasion and care.
To aid the computerisation of such tasks, active research is being undertaken
within the fields of Affective Computing (Scherer, et al., 2010; Picard, 2010),
and Social Robotics (Ge, 2007; Broekens, et al., 2009). While algorithms and
26
robots can now reproduce some aspects of human social interaction, the realtime recognition of natural human emotion remains a challenging problem, and
the ability to respond intelligently to such inputs is even more difficult. Even
simplified versions of typical social tasks prove difficult for computers, as is
the case in which social interaction is reduced to pure text. The social intelligence of algorithms is partly captured by the Turing test, examining the ability
of a machine to communicate indistinguishably from an actual human. Since
1990, the Loebner Prize, an annual Turing test competition, awards prizes to
textual chat programmes that are considered to be the most human-like. In
each competition, a human judge simultaneously holds computer-based textual
interactions with both an algorithm and a human. Based on the responses, the
judge is to distinguish between the two. Sophisticated algorithms have so far
failed to convince judges about their human resemblance. This is largely because there is much ‘common sense’ information possessed by humans, which
is difficult to articulate, that would need to be provided to algorithms if they are
to function in human social settings.
Whole brain emulation, the scanning, mapping and digitalising of a human brain, is one possible approach to achieving this, but is currently only a
theoretical technology. For brain emulation to become operational, additional
functional understanding is required to recognise what data is relevant, as well
as a roadmap of technologies needed to implement it. While such roadmaps exist, present implementation estimates, under certain assumptions, suggest that
whole brain emulation is unlikely to become operational within the next decade
or two (Sandberg and Bostrom, 2008). When or if they do, however, the employment impact is likely to be vast (Hanson, 2001).
Hence, in short, while sophisticated algorithms and developments in MR,
building upon with big data, now allow many non-routine tasks to be automated, occupa tions that involve complex perception and manipulation tasks,
creative intelligence tasks, and social intelligence tasks are unlikely to be substituted by computer capital over the next decade or two. The probability of an
occupation being automated can thus be described as a function of these task
characteristics. As suggested by Figure I, the low degree of social intelligence
required by a dishwasher makes this occupation more susceptible to computerisation than a public relation specialist, for example. We proceed to examining
the susceptibility of jobs to computerisation as a function of the above described
27
Public
Event
Relations
0 Planner
0
100
1 Court Clerk
0 Biologist
0
Social Intelligence
1
Fashion
Designer
Creativity
100
Probability of
Computerisation
Dishwasher
Probability of
Computerisation
Probability of
Computerisation
1
Telemarketer
Boilermaker
0
0
Surgeon
100
Perception and manipulation
F IGURE I. A sketch of how the probability of computerisation might vary as a function of
bottleneck variables.
non-susceptible task characteristics.
IV.
M EASURING
THE EMPLOYMENT IMPACT OF COMPUTERISATION
IV.A. Data sources and implementation strategy
To implement the above described methodology, we rely on O∗NET, an online
service developed for the US Department of Labor. The 2010 version of O∗NET
contains information on 903 detailed occupations, most of which correspond
closely to the Labor Department’s Standard Occupational Classification (SOC).
The O∗NET data was initially collected from labour market analysts, and has
since been regularly updated by surveys of each occupation’s worker population
and related experts, to provide up-to-date information on occupations as they
evolve over time. For our purposes, an important feature of O∗NET is that it
defines the key features of an occupation as a standardised and measurable set
of variables, but also provides open-ended descriptions of specific tasks to each
occupation. This allows us to: (a) objectively rank occupations according to
the mix of knowledge, skills, and abilities they require; and (b) subjectively
categorise them based on the variety of tasks they involve.
The close SOC correspondence of O∗NET allows us to link occupational
characteristics to 2010 Bureau of Labor Statistics (BLS) employment and wage
data. While the O∗NET occupational classification is somewhat more detailed,
distinguishing between Auditors and Accountants, for example, we aggregate
these occupations to correspond to the six-digit 2010 SOC system, for which
employment and wage figures are reported. To obtain unique O∗NET vari-
ables corresponding to the six-digit
SOC
28
classification, we used the mean of
the O∗NET aggregate. In addition, we exclude any six-digit SOC occupations
for which O∗NET data was missing.20 Doing so, we end up with a final dataset
consisting of 702 occupations.
To assess the employment impact of the described technological developments in ML, the ideal experiment would provide two identical autarkic
economies, one facing the expanding technological capabilities we observe,
and a secular decline in the price of computerisation, and the other not. By
comparison, it would be straightforward to examine how computerisation reshapes the occupational composition of the labour market. In the absence of
this experiment, the second preferred option would be to build on the implementation strategy of Autor, et al. (2003), and test a simple economic model
to predict how demand for workplace tasks responds to developments in ML
and MR technology. However, because our paper is forward-looking, in the
sense that most of the described technological developments are yet to be implemented across industries on a broader scale, this option was not available for
our purposes.
Instead, our implementation strategy builds on the literature examining the
offshoring of information-based tasks to foreign worksites, consisting of different methodologies to rank and categorise occupations according to their susceptibility to offshoring (Blinder, 2009; Jensen and Kletzer, 2005, 2010). The
common denominator for these studies is that they rely on O∗NET data in different ways. While Blinder (2009) eyeballed the O∗NET data on each occupation,
paying particular attention to the job description, tasks, and work activities, to
assign an admittedly subjective two-digit index number of offshorability to each
occupation, Jensen and Kletzer (2005) created a purely objective ranking based
on standardised and measurable O∗NET variables. Both approaches have obvi-
ous drawbacks. Subjective judgments are often not replicable and may result in
the researcher subconsciously rigging the data to conform to a certain set of be-
liefs. Objective rankings, on the other hand, are not subject to such drawbacks,
but are constrained by the reliability of the variables that are being used. At this
stage, it shall be noted that O∗NET data was not gathered to specifically mea20
The missing occupations consist of “All Other” titles, representing occupations with a
wide range of characteristics which do not fit into one of the detailed O∗NET-SOC occupations.
O ∗ NET data is not available for this type of title. We note that US employment for the 702
occupations we considered is 138.44 million. Hence our analysis excluded 4.628 million jobs,
equivalent to 3 percent of total employment.
29
sure the offshorability or automatability of jobs. Accordingly, Blinder (2009)
finds that past attempts to create objective offshorability rankings using O∗NET
data have yielded some questionable results, ranking lawyers and judges among
the most tradable occupations, while classifying occupations such as data entry
keyers, telephone operators, and billing clerks as virtually impossible to move
offshore.
To work around some of these drawbacks, we combine and build upon the
two described approaches. First, together with a group of ML researchers, we
subjectively hand-labelled 70 occupations, assigning 1 if automatable, and 0
if not. For our subjective assessments, we draw upon a workshop held at the
Oxford University Engineering Sciences Department, examining the automatability of a wide range of tasks. Our label assignments were based on eyeballing
the O∗NET tasks and job description of each occupation. This information is
particular to each occupation, as opposed to standardised across different jobs.
The hand-labelling of the occupations was made by answering the question
“Can the tasks of this job be sufficiently specified, conditional on the availability of big data, to be performed by state of the art computer-controlled equipment”. Thus, we only assigned a 1 to fully automatable occupations, where
we considered all tasks to be automatable. To the best of our knowledge, we
considered the possibility of task simplification, possibly allowing some currently non-automatable tasks to be automated. Labels were assigned only to
the occupations about which we were most confident.
Second, we use objective O∗NET variables corresponding to the defined
bottlenecks to computerisation. More specifically, we are interested in variables describing the level of perception and manipulation, creativity, and social
intelligence required to perform it. As reported in Table I, we identified nine
variables that describe these attributes. These variables were derived from the
O ∗NET survey, where the respondents are given multiple scales, with “importance” and “level” as the predominant pair. We rely on the “level” rating which
corresponds to specific examples about the capabilities required of computercontrolled equipment to perform the tasks of an occupation. For instance, in
relation to the attribute “Manual Dexterity”, low (level) corresponds to “Screw
a light bulb into a light socket”; medium (level) is exemplified by “Pack oranges in crates as quickly as possible”; high (level) is described as “Perform
open-heart surgery with surgical instruments”. This gives us an indication of
30
TABLE I. O∗NET variables that serve as indicators of bottlenecks to computerisation.
Computerisation
bottleneck
O ∗ NET
Perception
and
Manipulation
Finger
Dexterity
The ability to make precisely coordinated movements of
the fingers of one or both hands to grasp, manipulate, or
assemble very small objects.
Manual
Dexterity
The ability to quickly move your hand, your hand together
with your arm, or your two hands to grasp, manipulate, or
assemble objects.
Cramped Work Space,
Awkward Positions
How often does this job require working in cramped work
spaces that requires getting into awkward positions?
Originality
The ability to come up with unusual or clever ideas about
a given topic or situation, or to develop creative ways to
solve a problem.
Fine Arts
Knowledge of theory and techniques required to compose,
produce, and perform works of music, dance, visual arts,
drama, and sculpture.
Social
Perceptiveness
Being aware of others’ reactions and understanding why
they react as they do.
Negotiation
Bringing others together and trying to reconcile
differences.
Creative
Intelligence
Social
Intelligence
Variable
O ∗ NET
Description
Persuasion
Persuading others to change their minds or behavior.
Assisting and Caring for
Others
Providing personal assistance, medical attention, emotional support, or other personal care to others such as
coworkers, customers, or patients.
the level of “Manual Dexterity” computer-controlled equipment would require
to perform a specific occupation. An exception is the “Cramped work space”
variable, which measures the frequency of unstructured work.
Hence, in short, by hand-labelling occupations, we work around the issue
that O∗NET data was not gathered to specifically measure the automatability of
jobs in a similar manner to Blinder (2009). In addition, we mitigate some of the
subjective biases held by the researchers by using objective O∗NET variables to
correct potential hand-labelling errors. The fact that we label only 70 of the full
702 occupations, selecting those occupations whose computerisation label we
are highly confident about, further reduces the risk of subjective bias affecting
our analysis. To develop an algorithm appropriate for this task, we turn to
probabilistic classification.
31
IV.B. Classification method
We begin by examining the accuracy of our subjective assessments of the automatability of 702 occupations. For classification, we develop an algorithm
to provide the label probability given a previously unseen vector of variables.
In the terminology of classification, the O∗NET variables form a feature vec-
tor, denoted x ∈ R9 . O∗NET hence supplies a complete dataset of 702 such
feature vectors. A computerisable label is termed a class, denoted y ∈ {0, 1}.
For our problem, y = 1 (true) implies that we hand-labelled as computerisable
the occupation described by the associated nine O∗NET variables contained in
x ∈ R9 . Our training data is D = (X, y), where X ∈ R70×9 is a matrix of
variables and y ∈ {0, 1}70 gives the associated labels. This dataset contains
information about how y varies as a function of x: as a hypothetical example,
it may be the case that, for all occupations for which x1 > 50, y = 1. A
probabilistic classification algorithm exploits patterns existent in training data
to return the probability P (y∗ = 1 | x∗ , X, y) of a new, unlabelled, test datum
with features x∗ having class label y∗ = 1.
We achieve probabilistic classification by introducing a latent function
f : x 7→ R, known as a discriminant function. Given the value of the discriminant f∗ at a test point x∗ , we assume that the probability for the class label
is given by the logistic
(3)
P (y∗ = 1 | f∗ ) =
1
,
1 + exp(−f∗ )
and P (y∗ = 0 | f∗ ) = 1 − P (y∗ = 1 | f∗ ). For f∗ > 0, y∗ = 1 is more
probable than y∗ = 0. For our application, f can be thought of as a continuousvalued ‘automatability’ variable: the higher its value, the higher the probability
of computerisation.
We test three different models for the discriminant function, f , using the
best performing for our further analysis. Firstly, logistic (or logit) regression,
which adopts a linear model for f , f (x) = w ⊺ x, where the un-known weights
w are often inferred by maximising their probability in light of the training
data. This simple model necessarily implies a simple monotonic relationship
between features and the probability of the class taking a particular value.
Richer models are provided by Gaussian process classifiers (Rasmussen and
32
Williams, 2006). Such classifiers model the latent function f with a Gaussian
process (GP): a non-parametric probability distribution over functions.
A GP is defined as a distribution over the functions f : X → R such that the
distribution over the possible function values on any finite subset of X (such as
X) is multivariate Gaussian. For a function f (x), the prior distribution over its
values f on a subset x ⊂ X are completely specified by a covariance matrix K
(4)
p(f | K) = N (f ; 0, K) = √
1
det 2Ï€K
exp −
1 ⊺ −1
f K f .
2
The covariance matrix is generated by a covariance function κ : X × X 7→ R;
that is, K = κ(X, X). The GP model is expressed by the choice of κ; we consider the exponentiated quadratic (squared exponential) and rational quadratic.
Note that we have chosen a zero mean function, encoding the assumption that
P (y∗ = 1) = 12 sufficiently far from training data.
Given training data D, we use the GP to make predictions about the function
values f∗ at input x∗ . With this information, we have the predictive equations
(5)

p(f∗ | x∗ , D) = N f∗ ; m(f∗ | x∗ , D), V (f∗ | x∗ , D) ,
where
(6)
(7)
m(f∗ | x∗ , D) = K(x∗ , X)K(X, X)−1y
V (f∗ | x∗ , D) = K(x∗ , x∗ ) − K(x∗ , X)K(X, X)−1K(X, x∗ ) .
Inferring the label posterior p(y∗ | x∗ , D) is complicated by the non-Gaussian
form of the logistic (3). In order to effect inference, we use the approximate
Expectation Propagation algorithm (Minka, 2001).
We tested three Gaussian process classifiers using the GPML toolbox (Rasmussen and Nickisch, 2010) on our data, built around exponentiated quadratic,
rational quadratic and linear covariances. Note that the latter is equivalent to
logistic regression with a Gaussian prior taken on the weights w. To validate
these classifiers, we randomly selected a reduced training set of half the available data D; the remaining data formed a test set. On this test set, we evaluated
how closely the algorithm’s classifications matched the hand labels according
to two metrics (see e.g. Murphy (2012)): the area under the receiver operat33
TABLE II. Performance of various classifiers; best performances in bold.
classifier model
exponentiated quadratic
rational quadratic
linear (logit regression)
AUC
log-likelihood
0.894
0.893
0.827
−163.3
−163.7
−205.0
ing characteristic curve (AUC), which is equal to one for a perfect classifier,
and one half for a completely random classifier, and the log-likelihood, which
should ideally be high. This experiment was repeated for one hundred random
selections of training set, and the average results tabulated in Table II. The
exponentiated quadratic model returns (narrowly) the best performance of the
three (clearly outperforming the linear model corresponding to logistic regression), and was hence selected for the remainder of our testing. Note that its
AUC score of nearly 0.9 represents accurate classification: our algorithm successfully managed to reproduce our hand-labels specifying whether an occupation was computerisable. This means that our algorithm verified that our subjective judgements were systematically and consistently related to the O∗NET
variables.
Having validated our approach, we proceed to use classification to predict
the probability of computerisation for all 702 occupations. For this purpose,
we introduce a new label variable, z, denoting whether an occupation is truly
computerisable or not: note that this can be judged only once an occupation
is computerised, at some indeterminate point in the future. We take, again, a
logistic likelihood,
(8)
P (z∗ = 1 | f∗ ) =
1
.
1 + exp(−f∗ )
We implicitly assumed that our hand label, y, is a noise-corrupted version of
the unknown true label, z. Our motivation is that our hand-labels of computerisability must necessarily be treated as such noisy measurements. We thus
acknowledge that it is by no means certain that a job is computerisable given
our labelling. We define X∗ ∈ R702×9 as the matrix of O∗NET variables for all
702 occupations; this matrix represents our test features.
We perform a final experiment in which, given training data D, consisting
34
50
0
60
60
40
20
0.5
1
Probability of
Computerisation
100
50
0
50
0.5
1
Probability of
Computerisation
20
0
Cramped work space
Finger dexterity
40
60
40
20
0
0
0.5
1
Probability of
Computerisation
0
0.5
1
Probability of
Computerisation
0
0.5
1
Probability of
Computerisation
60
40
20
0.5
1
Probability of
Computerisation
80
60
1
0.5
Probability of
Computerisation
0
0
80
0
80
0
0
20
0.5
1
Probability of
Computerisation
Originality
100
40
0
0
Fine arts
Social perceptiveness
80
0
0
Manual dexterity
80
Negotiation
Persuasion
Assisting and
caring for others
100
0
0.5
1
Probability of
Computerisation
100
50
0
F IGURE II. The distribution of occupational variables as a function of probability of
computerisation; each occupation is a unique point.
35
of our 70 hand-labelled occupations, we aim to predict z ∗ for our test features
X∗ . This approach firstly allows us to use the features of the 70 occupations
about which we are most certain to predict for the remaining 632. Further, our
algorithm uses the trends and patterns it has learned from bulk data to correct
for what are likely to be mistaken labels. More precisely, the algorithm provides
a smoothly varying probabilistic assessment of automatability as a function of
the variables. For our Gaussian process classifier, this function is non-linear,
meaning that it flexibly adapts to the patterns inherent in the training data. Our
approach thus allows for more complex, non-linear, interactions between variables: for example, perhaps one variable is not of importance unless the value
of another variable is sufficiently large. We report P (z ∗ | X∗ , D) as the probability of computerisation henceforth (for a detailed probability ranking, see
Appendix). Figure II illustrates that this probability is non-linearly related to
the nine O∗NET variables selected.
V.
E MPLOYMENT
IN THE TWENTY- FIRST CENTURY
In this section, we examine the possible future extent of at-risk job computerisation, and related labour market outcomes. The task model predicts that recent
developments in ML will reduce aggregate demand for labour input in tasks
that can be routinised by means of pattern recognition, while increasing the demand for labour performing tasks that are not susceptible to computerisation.
However, we make no attempt to forecast future changes in the occupational
composition of the labour market. While the 2010-2020 BLS occupational employment projections predict US net employment growth across major occupations, based on historical staffing patterns, we speculate about technology that
is in only the early stages of development. This means that historical data on
the impact of the technological developments we observe is unavailable.21 We
therefore focus on the impact of computerisation on the mix of jobs that existed in 2010. Our analysis is thus limited to the substitution effect of future
computerisation.
Turning first to the expected employment impact, reported in Figure III, we
distinguish between high, medium and low risk occupations, depending on their
21
It shall be noted that the BLS projections are based on what can be referred to as changes
in normal technological progress, and not on any breakthrough technologies that may be seen
as conjectural.
36
Management, Business, and Financial
Computer, Engineering, and Science
Education, Legal, Community Service, Arts, and Media
Healthcare Practitioners and Technical
Service
Sales and Related
Office and Administrative Support
Farming, Fishing, and Forestry
Construction and Extraction
Installation, Maintenance, and Repair
Production
Transportation and Material Moving
400M
←−−− Low −−−→
33% Employment
←−−− Medium −−−→
19% Employment
←−−− High −−−→
47% Employment
Employment
300M
200M
100M
0M
0
0.2
0.6
0.4
Probability of Computerisation
0.8
1
F IGURE III. The distribution of BLS 2010 occupational employment over the probability of
computerisation, along with the share in low, medium and high probability categories. Note
that the total area under all curves is equal to total US employment.
37
probability of computerisation (thresholding at probabilities of 0.7 and 0.3).
According to our estimate, 47 percent of total US employment is in the high risk
category, meaning that associated occupations are potentially automatable over
some unspecified number of years, perhaps a decade or two. It shall be noted
that the probability axis can be seen as a rough timeline, where high probability occupations are likely to be substituted by computer capital relatively soon.
Over the next decades, the extent of computerisation will be determined by
the pace at which the above described engineering bottlenecks to automation
can be overcome. Seen from this perspective, our findings could be interpreted
as two waves of computerisation, separated by a “technological plateau”. In
the first wave, we find that most workers in transportation and logistics occupations, together with the bulk of office and administrative support workers,
and labour in production occupations, are likely to be substituted by computer
capital. As computerised cars are already being developed and the declining
cost of sensors makes augmenting vehicles with advanced sensors increasingly
cost-effective, the automation of transportation and logistics occupations is in
line with the technological developments documented in the literature. Furthermore, algorithms for big data are already rapidly entering domains reliant
upon storing or accessing information, making it equally intuitive that office
and administrative support occupations will be subject to computerisation. The
computerisation of production occupations simply suggests a continuation of a
trend that has been observed over the past decades, with industrial robots taking
on the routine tasks of most operatives in manufacturing. As industrial robots
are becoming more advanced, with enhanced senses and dexterity, they will be
able to perform a wider scope of non-routine manual tasks. From a technological capabilities point of view, the vast remainder of employment in production
occupations is thus likely to diminish over the next decades.
More surprising, at first sight, is that a substantial share of employment in
services, sales and construction occupations exhibit high probabilities of computerisation. Yet these findings are largely in line with recent documented technological developments. First, the market for personal and household service
robots is already growing by about 20 percent annually (MGI, 2013). As the
comparative advantage of human labour in tasks involving mobility and dexterity will diminish over time, the pace of labour substitution in service occupations is likely to increase even further. Second, while it seems counterintuitive
38
that sales occupations, which are likely to require a high degree of social intelligence, will be subject to a wave of computerisation in the near future, high
risk sales occupations include, for example, cashiers, counter and rental clerks,
and telemarketers. Although these occupations involve interactive tasks, they
do not necessarily require a high degree of social intelligence. Our model thus
seems to do well in distinguishing between individual occupations within occupational categories. Third, prefabrication will allow a growing share of construction work to be performed under controlled conditions in factories, which
partly eliminates task variability. This trend is likely to drive the computerisation of construction work.
In short, our findings suggest that recent developments in ML will put a substantial share of employment, across a wide range of occupations, at risk in the
near future. According to our estimates, however, this wave of automation will
be followed by a subsequent slowdown in computers for labour substitution,
due to persisting inhibiting engineering bottlenecks to computerisation. The
relatively slow pace of computerisation across the medium risk category of employment can thus partly be interpreted as a technological plateau, with incremental technological improvements successively enabling further labour substitution. More specifically, the computerisation of occupations in the medium
risk category will mainly depend on perception and manipulation challenges.
This is evident from Table III, showing that the “manual dexterity”, “finger
dexterity” and “cramped work space” variables exhibit relatively high values
in the medium risk category. Indeed, even with recent technological developments, allowing for more sophisticated pattern recognition, human labour will
still have a comparative advantage in tasks requiring more complex perception and manipulation. Yet with incremental technological improvements, the
comparative advantage of human labour in perception and manipulation tasks
could eventually diminish. This will require innovative task restructuring, improvements in ML approaches to perception challenges, and progress in robotic
dexterity to overcome manipulation problems related to variation between task
iterations and the handling of irregular objects. The gradual computerisation of
installation, maintenance, and repair occupations, which are largely confined to
the medium risk category, and require a high degree of perception and manipulation capabilities, is a manifestation of this observation.
Our model predicts that the second wave of computerisation will mainly
39
TABLE III. Distribution (mean and standard deviation) of values for each variable.
Variable
Assisting and caring for others
Persuasion
Negotiation
Social perceptiveness
Fine arts
Originality
Manual dexterity
Finger dexterity
Cramped work space
Probability of Computerisation
Low
Medium
High
48±20
48±7.1
44±7.6
51±7.9
12±20
51±6.5
22±18
36±10
19±15
41±17
35±9.8
33±9.3
41±7.4
3.5±12
35±12
34±15
39±10
37±26
34±10
32±7.8
30±8.9
37±5.5
1.3±5.5
32±5.6
36±14
40±10
31±20
depend on overcoming the engineering bottlenecks related to creative and social intelligence. As reported in Table III, the “fine arts”, “originality”, “negotiation”, “persuasion”, “social perceptiveness”, and “assisting and caring for
others”, variables, all exhibit relatively high values in the low risk category. By
contrast, we note that the “manual dexterity”, “finger dexterity” and “cramped
work space” variables take relatively low values. Hence, in short, generalist occupations requiring knowledge of human heuristics, and specialist occupations
involving the development of novel ideas and artifacts, are the least susceptible to computerisation. As a prototypical example of generalist work requiring a high degree of social intelligence, consider the O∗NET tasks reported for
chief executives, involving “conferring with board members, organization officials, or staff members to discuss issues, coordinate activities, or resolve problems”, and “negotiating or approving contracts or agreements.” Our predictions
are thus intuitive in that most management, business, and finance occupations,
which are intensive in generalist tasks requiring social intelligence, are largely
confined to the low risk category. The same is true of most occupations in
education, healthcare, as well as arts and media jobs. The O∗NET tasks of actors, for example, involve “performing humorous and serious interpretations of
emotions, actions, and situations, using body movements, facial expressions,
and gestures”, and “learning about characters in scripts and their relationships
to each other in order to develop role interpretations.” While these tasks are
very different from those of a chief executive, they equally require profound
40
Bachelor’s degree or better
Average median wage (USD)
80k
60k
40k
20k
0
0.5
1
60%
average
weighted by
employment
40%
unweighted
20%
0%
0
Probability of Computerisation
0.5
1
Probability of Computerisation
F IGURE IV. Wage and education level as a function of the probability of computerisation;
note that both plots share a legend.
knowledge of human heuristics, implying that a wide range of tasks, involving social intelligence, are unlikely to become subject to computerisation in the
near future.
The low susceptibility of engineering and science occupations to computerisation, on the other hand, is largely due to the high degree of creative intelligence they require. The O∗NET tasks of mathematicians, for example, involve
“developing new principles and new relationships between existing mathematical principles to advance mathematical science” and “conducting research to
extend mathematical knowledge in traditional areas, such as algebra, geometry,
probability, and logic.” Hence, whil…
Purchase answer to see full
attachment

  
error: Content is protected !!