Technological Forecasting & Social Change 114 (2017) 254–280

Contents lists available at ScienceDirect

Technological Forecasting & Social Change

The future of employment: How susceptible are jobs to computerisation?�

Carl Benedikt Freya,*, Michael A. Osborneb

aOxford Martin School, University of Oxford, Oxford OX1 1PT, United Kingdom bDepartment of Engineering Science, University of Oxford, Oxford OX1 3PJ, United Kingdom

A R T I C L E I N F O

Article history: Received 24 September 2015 Accepted 19 August 2016 Available online 29 September 2016

JEL classification: E24 J24 J31 J62 O33

Keywords: Occupational choice Technological change Wage inequality Employment Skill demand

A B S T R A C T

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 occupations probability of computerisation, wages and educational attainment.

© 2016 Published by Elsevier Inc.

1. Introduction

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 methodol- ogy to categorise occupations according to their susceptibility to computerisation.1 Second, we implement this methodology to esti- mate the probability of computerisation for 702 detailed occupa- tions, and examine expected impacts of future computerisation on US labour market outcomes.

� 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, Alex Flint, John Muellbauer, Vincent Mueller, Paul Newman, Seán Ó hÉigeartaigh, Anders Sandberg, Murray Shanahan, and Keith Woolcock for their excellent suggestions.

* Corresponding author. E-mail addresses: [email protected] (C. Frey), [email protected]

(M. Osborne). 1 We refer to computerisation as job automation by means of computer-controlled

equipment.

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 outrun- ning the pace at which we can find new uses for labour” (Keynes, 1933, p. 3). Indeed, over the past decades, computers have sub- stituted for a number of jobs, including the functions of book- keepers, 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 disagree- ment about the driving forces behind the persistently high unem- ployment 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

2 This view finds support in a recent survey by the McKinsey Global Institute (MGI), showing that 44% of firms which reduced their headcount since the financial crisis of 2008 had done so by means of automation (MGI, 2011).

http://dx.doi.org/10.1016/j.techfore.2016.08.019 0040-1625/© 2016 Published by Elsevier Inc.

C. Frey, M. Osborne / Technological Forecasting & Social Change 114 (2017) 254–280 255

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) empha- sise that the ongoing decline in manufacturing employment and the disappearance of other routine jobs is causing the current low rates of employment.3 In addition 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 sub- stantial 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 polarisa- tion, 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 tech- nological innovation is still increasing, with more sophisticated software technologies disrupting labour markets by making work- ers redundant. What is striking about the examples in their book is that computerisation is no longer confined to routine manufac- turing 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 employ- ment. The present study intends to bridge this gap in the literature. Although there are indeed existing useful frameworks for examining the impact of computers on the occupational employment composi- tion, they seem inadequate in explaining the impact of technological trends going beyond the computerisation of routine tasks. Semi- nal 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 devel- opments 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

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).

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 spe- cific 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 char- acteristics 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 occu- pational 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; Ingram and Neumann, 2006). Based on defined premises about what computers do, this literature examines the his- torical 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 (Brynjolf- sson 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 con- tent literature in a forward-looking manner. Furthermore, whereas this literature has largely focused on task measures from the Dictio- nary of Occupational 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 In particular, Ingram and Neumann (2006) use various DOT measurements to examine returns to different skills. Our analysis builds on their approach by classify- ing occupations according to their susceptibility to computerisation using O*NET data.

Second, our study relates to the literature examining the offshoring of information/based tasks to foreign worksites (Blinder, 2009; Blinder and Krueger, 2013; Jensen and Kletzer, 2005, 2010; Oldenski, 2012). 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 differ- ent occupations, Blinder (2009) estimates that 22 to 29% 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 off- shored: (a) the job must be performed at a specific work location; and (b) the job requires face-to-face personal communication. Naturally, the characteristics of occupations that can be offshored are differ- ent from the characteristics of occupations that can be automated. For example, the work of cashiers, which has largely been substi- tuted by self- service technology, must be performed at specific work location and requires face-to-face contact. The extent of computeri- sation 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 2, we review the literature on the historical relationship between technological progress and employment. Section 3 describes recent and expected future technological developments. In Section 4, we describe our methodology, and in Section 5, we examine the expected impact of these technological developments on labour market outcomes. Finally, in Section 6, we derive some conclusions.

2. A history of technological revolutions and employment

The concern over technological unemployment is hardly a recent phenomenon. Throughout history, the process of creative

4 Goos et al. (2009) provides a notable exception.

256 C. Frey, M. Osborne / Technological Forecasting & Social Change 114 (2017) 254–280

destruction, following technological inventions, has created enor- mous wealth, but also undesired disruptions. As stressed by Schum- peter (1962), it was not the lack of inventive ideas that 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 inven- tion, 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 inven- tion 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 persua- sively 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 tech- niques, 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 obso- lete 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 the Glorious Revolution of 1688, declined and lost most of its political clout (Nef, 1957, pp. 26 and 32). With Parliamen- tary 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 mech- anisation. 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 technologi- cal progress and deployed 12,000 men against the rioters (Mantoux, 2006, p. 403–408). 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 w