Cambridge Journal of Regions, Economy and Society 2020, 13, 153–173

doi:10.1093/cjres/rsz025
Advance Access publication 9 February 2020

Are machines stealing our jobs?

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

Andrea Gentilia, , Fabiano Compagnuccib, Mauro Gallegatic and
Enzo Valentinid
a
Faculty of Economics, Università degli Studi Internazionali di Roma, Via Cristoforo
Colombo, 200, 00147 Roma, Italy, [email protected]
b
Social Sciences, Gran Sasso Science Institute, Via Michele Iacobucci, 2, 67100 L’Aquila,
Italy, [email protected]
c
Department of Management, Università Politecnica delle Marche, Piazzale Martelli, 8,
60129 Ancona, Italy, [email protected]
d
Department of Political Science, Communication and International Relations, Università
degli Studi di Macerata, Via Don Minzoni, 2, 62100 Macerata, Italy, [email protected]

Received on February 28, 2019; editorial decision on November 27, 2019; accepted on December 17, 2019

This study aims to contribute empirical evidence to the debate about the future of work in
an increasingly robotised world. We implement a data-driven approach to study the techno-
logical transition in six leading Organisation for Economic Co-operation and Development
(OECD) countries. First, we perform a cross-country and cross-sector cluster analysis based
on the OECD-STAN database. Second, using the International Federation of Robotics
database, we bridge these results with those regarding the sectoral density of robots. We
show that the process of robotisation is industry- and country-sensitive. In the future, par-
ticipants in the political and academic debate may be split into optimists and pessimists
regarding the future of human labour; however, the two stances may not be contradictory.

Keywords: robotisation, labour dislocation, cluster analysis

JEL Classifications: E24, E66, J24

Introduction consensus that robots and artificial intelligence

(AI) will substantially affect all aspects of our
The purpose of this study is to offer an empir-
lives (Hägele et  al., 2016; Huang and Rust,
ical contribution to the debate on the effects of
2018; Makridakis, 2017; OECD, 2018c), espe-
robots on productivity, employment and wages,
cially employment and wages. However, the
by investigating the density of robots and la-
predictions of and reactions to their effects on
bour dislocation in the economic sectors of
societies and economies have been mixed, ran-
six leading countries of the Organisation for
ging from utopian to dystopian visions of the
Economic Co-operation and Development
future. According to Dosi and Virgillito (2019,
(OECD). In the debate on the future of labour,
4): ‘New technologies may herald an economics
which has raised significant interest in both the
of hope, with work for all and equitable social
academic and political fields, there is general

© Crown copyright 2020.

This article contains public sector information licensed under the Open Government Licence v3.0
(http://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/).
Gentili et al.

inclusion, or conversely, mass unemployment, of robots and the pace of their technological
mass inequality and social exclusion, leading to growth have been industry-sensitive (Bessen,
a “re-feudalisation” of Western societies’.1 The 2017; Camagni, 2017; Compagnucci et al., 2019;

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

existence of radically opposite views of robot- Mandfield, 1989). Until now, the introduction of
isation and its effects is not surprising because robots has only partially concerned the service
of the unprecedented complexity and perva- sector. The main effect is on the manufacturing
siveness of this process (Korinek and Stiglitz, sector (International Federation of Robotics,
2017), which is leading the transition to the 2018), especially high- and medium-tech
Fourth Industrial Revolution (Brynjolfsson manufacturing (Piva and Vivarelli, 2018), which
et al., 2017). has the highest readiness for investment in
The European Commission (EC, 2019), for innovation. In splitting the production pro-
instance, assumes that two scenarios are likely. cess into different phases, automatable tasks
The first scenario is pessimistic. Robotisation are performed by machines, whereas non-
and AI-based machine learning will improve automatable tasks (that is, so-called residual
the prediction and decision-making capability activities) are carried out by humans (Decker,
of machines, leading to the prevalence of the 2017). The sectoral risk of job automation was
substitution effect on the complementarity ef- confirmed empirically by the OECD (2018a,
fect between workers and machines and, con- 2018b), showing that, at least since 2011, re-
sequently, disruptive outcomes in terms of gions that are specialised in tradable services,
job losses and the disqualification of certain which are also highly urbanised and have highly
typologies of workers. This effect might prove educated human capital, are the least exposed
to be so important that it causes medium and to this threat.
even long-term negative effects on employ- However, the rapid increase in the number
ment and wages. In contrast, the second scen- and types of tasks that a robot can perform, ran-
ario is optimistic, suggesting that robots and AI ging from manual and routine cognitive tasks to
will increase labour productivity, thanks to the non-routine manual and cognitive tasks, and the
complementarity between machines and hu- fact that not only agriculture and manufacturing
mans, the latter of which will benefit from the (that is, tangible production) but also the service
ability of machines to do routine tasks because sector (that is, non-tangible production) are sub-
of the freedom to perform non-routinised so- ject to robotisation and AI, has increased con-
cial and intellectual jobs. Moreover, workers cerns about the future of work (Pew Research
in jobs that complement technologies will ex- Center, 2017). The ongoing wave of robotisa-
perience substantial wage increases. However, tion could affect routinised and low-skilled jobs
workers that will be substituted by machines (OECD, 2018b), and it might also displace well-
will experience diminished compensation and paid and skilled jobs performed by highly edu-
work opportunities. In addition, sectors where cated workers in the advanced services sector.
labour complements technology could absorb Big data and improvements in programming
routine workers who leave mainly from the and algorithm applications have already be-
manufacturing sector2 because of a higher ‘job come the standard in the financial, credit and
multiplier’3 (Moretti, 2012), which ultimately insurance sectors. More importantly, the new
would lead to a higher level of long-run equi- frontier of robotics is focussed on the develop-
librium in terms of employment, compensation ment of robots that can expand the physical and
and gross domestic product (GDP). mental abilities of human beings in numerous
While there is no consensus on the effects tasks, ranging from healthcare and medical
of robotisation, there is evidence that the use intervention to defence and space exploration.

154
 Are machines stealing our jobs?

This cutting-edge technology includes the de- robots on our societies. Unlike other contri-
velopment of social robots and their applica- butions based on IFR data (Acemoglu and
tion to human–robot interactions (Breazeal Restrepo, 2018; Graetz and Michaels, 2017),

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

et  al., 2008), such as supporting the process of which investigated the causal relationship be-
learning and training in work and education tween labour dislocation and robotisation over
and helping patients in the execution and moni- a longer period, in this work no hypotheses are
toring of physical rehabilitation and cognitive stated before we conduct the database analysis.
exercises (González et al., 2017). However, although our approach has allowed
If it is true that robotisation will produce dif- us to achieve results that are not biased by a
ferent effects depending on the time and the priori assumptions and restrictions, it obliges
sector (Bessen, 2017), it is also true that these us to work with a limited database, which pre-
effects will be country- and region-sensitive, vented the investigation of causal relationships.4
because the latter have different economic The article proceeds with a literature review,
specialisations and are in different stages of de- followed by a section describing the data and
velopment. A more or less knowledge-intensive the methodology before presenting the results
service-based economy or a more or less of the cluster analysis. The fourth section dis-
high-tech manufacturing production system, as cusses the role of robotisation in enhancing
well as the institutional framework and educa- productivity and causing labour dislocation,
tion level (Acemoglu, 2002, 2015), could sub- and is followed by a brief conclusion.
stantially affect the adoption of technology.
Furthermore, this process could further exacer-
bate spatial inequalities (Giaoutzi and Stratigea, Technological change in the time
1991; Sujarwoto and Tampubolon, 2016). of robots and AI: is it different
Without neglecting the importance of insti- this time?
tutional factors, to identify the economic sec- Technology has always been a key topic in eco-
tors and countries that have suffered severely nomic research because of its effects on the
from labour dislocation in the last two decades performances of single firms, industries and
and to determine whether there are significant countries as well as on workers, human capital
differences in the density of robots and their and the income distribution between labour
growth rate among country–sector pairs, in and capital. Since the late 18th century, the
this study we focus on a set of variables (that main concern regarding technological change
is, productivity, value added, relative price, hour has been its capacity to substitute human la-
worked, hourly wages and total wages) drawn bour by machines, possibly leading to job losses
from the OECD-Structural Analysis (STAN) or so-called ‘technological unemployment’
database. We perform a cluster analysis using (Bennion, 1943). Although most scholars agree
these variables and provide a taxonomy of on the possibility that temporary unemploy-
the different clusters resulting from the ana- ment would follow the introduction of new
lysis. We then relate these clusters to the on- labour-saving technologies, there is no con-
going robotisation process using data collected sensus about their long-term effects (Mokyr
from the International Federation of Robotics et al., 2015).
(IFR). Our analysis supports the hypothesis As Brynjolfsson et al. (2017, 33) pointed out,
that the technological transition is industry- ‘there are plenty of both optimists and pessim-
and country-sensitive. ists about technology and growth’. According
In using a data-driven approach, we aim to the pessimistic scenario, the main concerns
to contribute to the debate on the effects of have been the decreased importance of human

155
Gentili et al.

labour compared with automated tasks, and the transformations during the last three decades:
impossibility of absorbing the surplus work- the ongoing process of globalisation and polar-
force in alternative employment opportunities. isation of manufacturing activities in China; the

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

The labour-saving effects triggered by new rising importance of the knowledge economy
technologies have been predicted to be higher and the related dynamics in terms of wage po-
than the creation of new jobs (Arntz et al., 2016; larisation and workforce casualisation; the
Frey and Osborne, 2017; Leontief, 1952). Based increasing decoupling of productivity trends on
on the past two centuries, it may be agreed that the one hand and wage, employment and com-
the doomsayer’s predictions have not been ful- pensation trends on the other. After the 2007
filled. Mokyr (2002), for example, did not find crisis, notwithstanding the report that ‘labour
evidence that technological unemployment oc- markets are back to pre-crisis levels in terms of
curred on a large scale during the first industrial job quantity, with only a few notable exceptions’
revolution, when mechanisation replaced only OECD (2018a), this trend was accompanied by
a few human activities. Similarly, Saint-Paul a higher proportion of poor-quality jobs (that is,
(2017), in discussing the introduction of the casual and precarious jobs), which limit the wage
assembly line in the automotive sector in the growth (Chandrasekhar, 2018) of the so-called
early 19th century, found that it complemented ‘working poor’, that is, people with low-paid jobs,
low-skilled workers while it substituted highly whose incomes fall below the poverty line.
skilled craftsmen, finally producing positive ef- Considering this context, what impact do we
fects in the long run. According to Piketty and expect from robots and AI, and which perspec-
Zucman (2014), during Les Trente Glorieuses5 tive appears to be closer to the empirical evi-
(the Glorious Thirty) years, productivity, em- dence? Answering this question is particularly
ployment and wages increased, triggering posi- challenging, not only because of the theoretical
tive effects on well-being and the distribution complexity of the issue but also because of the
of wealth, finally reducing the gap between the lack of affordable and reliable data, which pre-
upper and lower classes. vents the increase in the quantity and quality of
Despite the evidence that improvements in empirical studies.
technology do not necessarily result in techno- Among the studies that assessed the ef-
logical unemployment (at least not in the long fects of robots on productivity, Kromann
run) based on what has happened so far, why et al. (2011) assumed that the stock of indus-
are robotisation and AI again splitting the aca- trial robots affected the level of automation.
demic and political debate between pessimists In their study, the results of a cross-country
and optimists? Is it perhaps different this time? and cross-industry analyses of the European
Frank et al. (2019, 1) noted, ‘rising automation Union level analysis of capital (K), labour
is happening in a period of growing economic (L), energy (E), materials (M) and service
inequality, raising fears of mass technological (S) inputs (EU-KLEMS) and IFR data be-
unemployment and a renewed call for policy tween 2004 and 2007 showed that automa-
efforts to address the consequences of techno- tion increased labour productivity, causing
logical change’, which could lead to the rise of negative short-run and positive long-run ef-
the so-called ‘useless class’ (Harari, 2014) of hu- fects on employment. Similarly, the European
mans who are not able to work because their Commission Report on Robotics and
professions have become obsolete. Dosi and Employment (EC, 2016) showed that the use
Virgillito (2019) argued that robotisation and of industrial robots allowed companies to
AI have intervened in the socioeconomic con- significantly increase their level of product-
text, which has experienced radical, interlinked ivity. Graetz and Michaels (2015) considered

156
 Are machines stealing our jobs?

14 industries in 17 countries from 1993 to increasing share of obsolete investment in

2007 and performed an ordinary linear re- total investments;
gression (OLS) and two-stage least squares 2) a marked lag between the implementation of

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

(2SLS) analyses on EU-KLEMS and IFR a technology and the occurrence of its effects;
data. They found that the pairs of country–in- 3 ) measurement problems.
dustry that experienced the highest level of
robot densification also showed the highest There are conflicting pieces of evidence re-
gains in labour productivity. However, ‘larger garding not only the relation between robotisa-
increases in robot density translated into in- tion and AI on the one hand and productivity
creasingly small gains in productivity, sug- on the other hand, but also their possible ef-
gesting that there are some congestion effects fects on employment and wages. The review
(or diminishing marginal gains)’ (Graetz and of the related literature reveals that the po-
Michaels, 2015, 3). Finally, Dauth et al. (2018), tential outcomes depend on the features of
using data on German labour markets be- and interplay among the industries involved.
tween 1994 and 2014, found that regions that According to Vivarelli (2014), Acemoglu
were more exposed to automation showed sig- (2002) and Acemoglu and Restrepo (2018),
nificant increases in labour productivity. The the labour-saving effects of technological im-
empirical evidence found in these previous provement can be compensated by the fol-
studies was supported by predictions by the lowing dynamics in different contexts: (i) at the
Centre for Economics and Business Research intra-industry level, when the increasing prod-
(2017) and the Boston Consulting Group uctivity in a given sector spurs the growth of
(2015), according to which productivity was its output allowing increased labour demand
expected to substantially increase because of for non-automated tasks (and jobs) within the
autonomation in the coming decades. same sector, and (ii) at the cross-industry level,
Nonetheless, although AI and robotisation when the increasing productivity in one sector
have been expected to have a major impact positively affects the employment in other sec-
on productivity, recent data indicate a marked tors through the price effect, wage effect, new
and indisputable slowdown in global product- industries effect and new investment effect. The
ivity, especially since the global financial crisis price effect occurs whenever non-automated
(United Nations, 2017). Surprisingly, and para- sectors expand their output and reduce their
doxically, this slowdown ‘appears orthogonal cost production in response to the fall in rela-
to recent technological advances’ (Goldin tive prices in the automated sectors, which then
et  al., 2018, 1). The productivity slowdown increases their labour demand. The wage effect
was confirmed by the OECD (2019), which can positively affect both the automated and
underlined the existence of substantial differ- the non-automated sectors. In the first case,
ences among countries, industries and firms in the rise in workers’ wages in innovative sec-
adopting technological improvement, as well tors positively affects the demand for goods
as the ways these gaps affect productivity. and services, eventually causing an expansion.
A review of the possible causes of this ongoing In the second case, decreasing wages in non-
trend was provided by Goldin et  al. (2018) automated sectors leads to a reduction in their
and Brynjolfsson et al. (2017), who identified production costs and, potentially, their expan-
manifold causes: sion. The new industries effect arises from the
increase in the use of robots, which, to be pro-
1) the way in which technological change
duced, require new industrial plants and work-
is taking place is characterised by an forces. These new occupations have specific

157
Gentili et al.

needs that are oriented toward new technolo- will fall, leading to the expulsion of the labour
gies and new products, thus reinforcing the force. In the case of barriers to labour mobility,
growth of the whole sector. Finally, the new in- the workforce surplus remains trapped in the

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

vestment effect refers to evidence that, because distinctive sector, causing a reduction in dispos-
the accumulation of capital is the engine of able income and, depending on the size of the
economic growth, and technological improve- sector, a fall in aggregate demand, which could
ments augment capital, increases in technolo- spread throughout the entire economy. Delli
gies are reflected in increases in labour. Gatti et  al. (2012, 375)  noted, ‘Nowadays […]
However, no condition ensures that one (or falling incomes in manufacturing yield a lack of
many) of the above conditions hold. As argued demand for goods produced in the rest of the
in Cirillo et al. (2015), at the industry level la- economy, namely the service sector’.
bour demand can expand if and only if the Finally, there may be labour displacement in
output growth is larger than the productivity advancing industries even in the presence of
growth, the occurrence of which, however, is high elasticity. However, this result is ‘incon-
not obvious. First, there can be constraints on sistent with models of structural change that
aggregate demand growth, which was pointed assume an underlying Cobb–Douglas produc-
out by Vivarelli and Pianta (2000) with re- tion structure in each industry’, as pointed out
spect to stability rules in the European Union. in Autor and Salomons (2018, 7). Moreover,
Second, in the automation sectors, job creation even when the condition of the expansion of
depends on the sectoral elasticity of demand, aggregate demand is met, we may observe la-
which may vary over time and across sectors bour dislocation due to constraints on labour
(Bessen, 2018). Furthermore, Bessen (2017) mobility. As noted by Vivarelli (2014), techno-
found that new technologies should have a logical bias, which consists of a mismatch be-
positive effect on employment if they improve tween workers’ skills and those required by
productivity in markets where there is a large new occupations in technologically advanced
amount of unmet demand. In the context of ro- sectors (Acemoglu, 2002), could cause labour
botics and automation, he suggested that new dislocation independently from labour de-
computer technology is associated with em- mand. Acemoglu and Restrepo (2018) stress
ployment declines in manufacturing, where de- that in the USA, robotisation caused labour
mand has generally been met but is correlated dislocation mainly in sectors affected by auto-
with employment growth in less-saturated non- mation, both in terms of labour utilisation (the
manufacturing industries. negative impact of which could be absorbed by
Third, in a fast-growing industry, the aggre- other non-automating sectors) and of labour
gate demand could be inelastic, eventually shares. In Graetz and Michaels (2017), although
resulting in a reduction in labour demand, as increasing productivity due to robotisation did
Delli Gatti et  al. (2012) explained, based on not appear to affect post-crisis recoveries of
the extended crisis theory (ECT). The authors world-leading economies (except the USA),
showed that a persistent structural problem it had, a negative impact on labour share. As
could arise when a large but distinctive sector Felten et al. (2008) pointed out, if it is true that
(for example, agriculture during the Great AI may boost growth, the consequences for la-
Depression, but manufacturing nowadays) of bour are less clear according to the preponder-
an economy faces an uneven increase in prod- ance of complementarity or the substitutability
uctivity because of technological improve- effect between humans and machines.
ments. If this sector faces an inelastic demand To conclude, Acemoglu (2002), the
curve, the relative price of the goods produced OECD (2012), Gries and Naudé (2018) and

158
 Are machines stealing our jobs?

Brynjolfsson and McAfee (2014) focussed Empirical analysis

on the role played by technological bias in
The empirical analysis is aimed at identifying
shaping our societies. In particular, Acemoglu
the existence of common patterns among the

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

(2002) stressed the importance of techno-
economic sectors in different countries and
logical bias in causing labour dislocation,
their relationship with the robotisation pro-
while Gries and Naudé (2018, 23) highlighted
cess. We perform a cluster analysis, which is a
the key role played by technology in shaping
multivariate statistical technique that splits
the labour market: ‘if labour income does not
a set of uncategorised data into a k number
profit from the economic gains generated
of groups (or clusters) based on their simi-
by progress in AI [and robotisation], con-
larity. Observations belonging to the same
sumption may stagnate and restrict growth’.
cluster are more similar to each other than to
In their model, stagnation of labour income
those included in other clusters. Cluster ana-
de facto translates to stagnant aggregate de-
lysis is based on an iterative algorithm that,
mand. Finally, Brynjolfsson and McAfee
starting in k initial cluster centres, minimises
(2014) noted that, beginning in the 1980s, US
the Euclidean distance within each cluster be-
productivity, median income and employ-
tween its mean or medians (k-clustering) and
ment, which had grown at the same pace after
its observations while maximising the distance
the Second World War, began to follow dif-
in terms of means or medians among the adja-
ferent trends; they termed this phenomenon
cent clusters.
the ‘Great Decoupling’. While productivity
The data-driven approach, which makes it
continued to increase, the growth of median
possible to group a large number of observa-
income slowed down and became negative
tions based on several variables, has been used
during the last two decades. Employment fol-
in a wide range of studies from cross-sectional
lowed the same trend, becoming almost flat be-
to time-series and panel data analyses. Repkine
tween 2000 and 2007 when the economy was
(2012), for instance, identified groups of East
still expanding. They showed that the ‘Great
Asian countries that were similar in economic
Decoupling’ mainly depended on the uneven
terms, whereas Green and Henseke (2016)
distribution of salaries. College graduate in-
performed a conventional cluster k-medians
comes have constantly grown since the 1980s,
analysis to split occupations into ‘graduate’
while the rest of the US labour force has fol-
and ‘non-graduate’ jobs. In addition, Nasri and
lowed a flat income trend since the 1980s.
Zhang (2018) used a multi-dimensional cluster
Workers may have been replaced by machines
approach to measure the spatial structure of
in automatising sectors, and because of lack
US metropolitan areas, and Kontsevaya (2017)
of the required skills, they are constrained to
applied a k-mean cluster approach to estimate
unemployment and/or to the reallocation to
the efficiency of the utilisation of subsidies
non-automatising, less knowledge-intensive
to investments in agriculture in the Russian
and less well-paid sectors (Compagnucci
Federation. Theis and Weihs (2000) introduced
et al., 2018). In any case, this labour realloca-
a variant of cluster analysis, which is known as
tion reduced the labour share, strengthening
fuzzy-cluster analysis, to identify the number
the fall in aggregate demand and potentially
of distinct stages in Germany’s business cycles.
bringing on long-run stagnation or crisis. In
Finally, according to Beck  et  al. (2017, 12),6
the following sections, we determine whether
‘k-means clustering analysis […] has been iden-
this negative outlook is compatible with the
tified as an appropriate technique, as it allows
empirical evidence and, eventually, the role of
the investigation of characteristics which may,
robots in this process.

159
Gentili et al.

in combination, differentiate between sharing productivity dynamics. We labelled as A  the

and non-sharing economy businesses’. clusters that showed an increase in product-
In our analysis, we used the k-median ivity and a relative price reduction (which was

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

method rather than the k-mean method be- expected, since automation reduces the costs
cause the former is less sensitive to outliers, of production). On the contrary, we labelled
whose presence affected our database. The as B the clusters showing a flat productivity
‘cluster’ approach allowed us to group the pairs trend and growing relative prices. In addition,
sector i and country j7 (that is, our observations) because we were interested in labour disloca-
according to the trends in some selected vari- tion, we focussed on wage dynamics using one
ables (that is, productivity, wages, prices and or more positive or negative signs (+, ++, +++
value added over time). After identifying clus- and -, --, ---) to represent their intensity. Thus, if
ters with homogeneous behaviours, we created a sector behaved accordingly with the positive
a taxonomy related to the specific behaviours vision presented in the literature (for example,
of our subjects in the cluster according to the if improvements in technology were mirrored
economic variables used. both in sectoral employment and compensa-
The analysis was based on the OECD-STAN tion growth), it was included in an A++ cluster,
and IFR databases. STAN provides economic which thus experiences increasing productivity
information over the period 2000–2015 for (A), increasing labour utilisation (+) and com-
OECD countries, and IFR supply data about pensation (+). If a sector showed flat prod-
the stock and growth rate of robots by sector uctivity, flat labour utilisation and decreasing
and country. We chose to analyse six OECD compensations, it was included in a B- cluster
leading countries: Denmark, France, Germany, and so on.
Italy, the UK and the USA,8 for the following The identification of clusters A and B helped
reasons: (a) they represent different models us to corroborate Bessen’s (2017) theoret-
of development (this characteristic is particu- ical hypotheses of labour dislocation, as well
larly interesting because, as discussed above, as those underpinning the ECT (Delli Gatti
both space and industry specialisation matter); et al., 2012). The relative dimension of cluster
and (b) comparable data for these countries A, in which the labour demand may increase or
are available from both databases (Meliciani, decrease, and the capacity of labour to be re-
2001). The analysis covered 21 economic sec- located in sectors B+, B- or in unemployment,
tors, which, in addition to the six considered were crucial factors in determining whether
countries, resulted in 126 statistical units. The the optimistic or the pessimistic vision was the
total value added (VALK), relative prices most plausible.
proxied using sectoral and GDP deflators The proposed approach to identifying the
(VALP), hours worked (HRSE), hourly prod- country–sector pairs falling in clusters A and B
uctivity (VALK/HRSE), total wages (WAGE) had two main advantages. First, as already men-
and hourly wages (WAGE/HRSE, deflated tioned, it allowed us to avoid bias, which could
using the sectoral deflator) were drawn from arise from applying predefined classifications
STAN. Robot density and the average growth to the data. Depending on the different institu-
rate of robots (discussed in the next section) tional settings and economic specialisations, the
were drawn from IFR. composition of clusters A  and B varied from
A key role in determining the existence of la- country to country. Second, the cluster approach
bour dislocation due to technological improve- was more useful in linking these dynamics with
ments is played by productivity. Therefore, the robotisation process. Because robots are
we classified our clusters according to their a developing technology that requires large

160
 Are machines stealing our jobs?

investment, arguably they do not spread simul- Although the log (WSS) and the PRE sup-
taneously throughout each country–sector. On ported this value, they did not preclude that
the contrary, it is more likely that they spread we could also consider the k = 11 solution. In

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

faster in the sectors in which countries invest examining the cluster composition, however,
heavily, which may differ considerably between we found that three additional clusters re-
countries. lated to the k = 11 solution were composed of
In analysing the data, we focussed on vari- one single sector; therefore, we chose the k = 8
ations instead of levels because the countries solution. However, even in the eight-cluster
and sectors differed in terms of size. Specifically, solution, we observed three single sector
we considered the index number of each vari- clusters: information and communication in
able with a base value of 100 in 2000. We per- Denmark; electrical, electronic and optical
formed the cluster analyses using the Stata equipment in France and the USA.
cluster kmedians command. In Figure 1, which Table 1 shows the descriptive statistics of the
shows the within sum of square (WSS) matrix, five clusters composed of more than one sector,
its logarithm, the η 2 coefficient, and the propor- and Table 2 reports the distribution in clusters
tional reduction of error (PRE), the plots in- of each sector in each country. Subsequently,
dicate that the optimal number of clusters was we labelled each cluster according to the char-
eight (for technical details, see Makles, 2012). acteristics considered. As expected, although
At k  =  8 (being k the number of clusters), the analysis was more complex, the choice of
there was a kink in the WSS corresponding avoiding predefined classifications appeared to
to a η82reduction in the WSS of almost 80%. lead to more realistic results.
800000

13.5
600000

13
log(WSS)
12.5
WSS
400000

12
200000

11.5
11
0

0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
k k
.4
1
.8

.3
.6

.2
PRE
2

.4

.1
.2

0
−.1
0

0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
k k

Figure 1.  WSS, log (WSS), η2 and PRE for all K cluster solutions.
Note: Authors’ elaboration from the STAN database.

161
Gentili et al.

Table 1.  Cluster (k-median) results

Cluster Productivity Value added Relative price Hourly wage Hours worked Total wages

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

A++ Mean 170.45 177.50 71.57 167.89 106.24 175.81
Median 166.87 182.92 67.27 174.58 109.62 174.33
N 9 9 9 9 9 9
A+ Mean 141.85 111.29 86.93 140.57 79.66 110.87
Median 133.92 108.34 88.90 140.24 77.04 111.10
N 19 19 19 19 19 19
A- Mean 143.50 73.54 89.92 140.70 52.40 72.55
Median 135.50 74.73 89.48 139.70 57.93 73.27
N 17 17 17 17 17 17
B+ Mean 104.41 125.16 104.25 109.5718 122.4583 132.5686
Median 110.70 125.13 103.18 110.4702 118.2017 132.127
N 38 38 38 38 38 38
B- Mean 103.21 100.03 115.36 97.42 98.12 94.67
Median 100.93 98.06 112.78 93.75 97.17 97.05
N 40 40 40 40 40 40

Note: Authors’ elaboration from the STAN database. Mean and median values are computed on the ‘n’ country sectors in
the cluster.

Specifically, the dynamics of each cluster was Other question could be raised regarding
as follows: ICT services and transport and equipment
in the UK. The prices of ICT services de-
• A++: increasing productivity, increasing
creased because of the strong international
value added, decreasing prices, increasing
competition during the period of analysis.
hourly wages, increasing hours worked
Productivity, as well as demand and added
and increasing total wages. This cluster
value, increased because of continuous
included knowledge-intensive and
technological advances, whereas hours
high-tech activities, such as Information
worked remained stable. Here again, as the
and Communications Technology (ICT)
increase in hourly wages signalled, workers
(in every country but Denmark), transport
benefited from the increased productivity.
equipment (in Germany and the UK), and
In these sectors, the demand for skilled
electrical, electronic and optical equip-
workers is often higher than the supply,
ment (in Germany). Some comments could
which leads to an adequate compensation
be raised about the presence of transport
level. All these considerations applied to
equipment, which is generally considered
transport equipment in the UK because
a mature market (that is, increasing prod-
of the relative importance of the manu-
uctivity, decreasing prices and stagnating
facture of aircraft and spacecraft in that
demand). Regarding Germany, the hours
sector. The manufacture of aircraft and
worked did not decrease because of the
spacecraft is a high-tech sector in which
increasing competitiveness of German
the UK plays a leading role. Therefore, this
automotive production9 in the global
cluster showed the most positive return to
market. However, the increase in ‘hourly
increased productivity. Increasing prod-
wage’ showed that workers benefited from
uctivity was associated with an increase
the increased productivity, which was
in employment as in the case of elasticity
confirmed by the increase in total wages.
of demand larger than one (Bessen, 2017).

162
 Are machines stealing our jobs?

Table 2.  Country–sector distribution of clusters

Denmark France Germany Italy UK USA

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

Manufacturing and construction
 Construction B- B- B- B- B+ B-
  Electrical, electronic and optical equipment A+ Out A++ A+ A- Out
  Food products, beverages and tobacco A- B+ B- B- A+ B-
 Furniture; other manufacturing; repair and installation A+ A+ B- B- A- A+
of machinery and equipment
  Textiles, wearing apparel, leather and related products A- A- A- A- A- A-
  Transport equipment A- B- A++ A+ A++ A+
  Wood and paper products, and printing A- A+ A+ A- A- A-
 Basic metals and fabricated metal products; machinery and B- A+ B- A+ A- B-
equipment
 Plastic and chemical products, glass, ceramics, stone, mineral A+ A+ A+ A- A- B-
products
Services
  Accommodation and food service activities B+ B+ B- B+ B+ B+
  Arts, entertainment and recreation B- B+ B- B+ B+ B+
 Education B- B- B- B- B- B-
  Financial and insurance activities B+ B+ B- B- B- B+
  Human health and social work activities B+ B+ B+ B- B+ B+
  Information and communication Out A++ A++ A++ A++ A++
  Other service activities B- A+ B- B+ B+ B-
  Professional, scientific and technical B+ B+ B+ B+ A++ B+
  Public administration and defence; compulsory social security B- A+ A+ B- B- B+
  Real estate activities B+ B+ B+ B- B+ B+
  Transportation and storage B- B- B+ B- B- B-
 Wholesale and retail trade, repair of motor vehicles and B+ B+ A+ B+ B+ B+
motorcycles

Note: Authors’ elaboration from STAN database.

This cluster mainly supported the positive may have caused increasing wage inequality
vision of the structural effects of robots and unemployment.
and AI. • A-: Increasing productivity, decreasing
• A+: Increasing productivity, increasing value added, decreasing prices, increasing
value added, decreasing prices, increasing hourly wages, substantially decreasing
hourly wages, decreasing hours worked and hours worked and decreasing total wages.
increasing total wages. Despite the reduction All sectors included in this cluster be-
in the number of hours worked, this cluster longed to manufacturing. It is worth
fully recovered the loss in aggregate de- noting that the textile sector in all coun-
mand by increasing wages. This cluster was tries was included in this cluster and
compatible with skill-biased technological that the UK had the highest number of
change theory (Autor et al., 2003; Autor and manufacturing sectors in this cluster, fol-
Dorn, 2013). On the one hand, technological lowed by Denmark and Italy. A- was com-
improvement and capital raised the demand patible with sectors where the elasticity of
for highly skilled workers, and consequently, demand was lower than 1 (Bessen, 2017),
their wage premium; on the other hand, they which meant that increasing productivity

163
Gentili et al.

generated negative structural effects. It be- the austerity policies following the 2007
haved as a fast-growing sector facing the crisis. Financial and insurance activities,
inelastic demand of ECT; hence, its labour on the contrary, were affected negatively

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

utilisation at the end of the period was al- mainly by ICT diffusion, which triggered a
most halved. shift towards Internet-based services (as in
• B+: flat productivity, increasing value the case of mobile banking), thus reducing
added, increasing prices, increasing hourly the workforce in those sectors.
wages, increasing hours worked and
increasing total wages. It included only two Figure 2 shows the dynamics of each cluster be-
industrial sectors (construction in the UK tween 2000 and 2015 in terms of hours worked
and food and beverages in France) and and wages. Specifically, the relative dimension
about half of the economic activities in the of each cluster in terms of the variable under
service sector. B+ behaved similarly to the investigation is reported on the left axis, and
tertiary sector in Baumol (1967) and the the right axis shows the total hours (left graph),
stagnant (slowly growing) sector in ECT. total wages and labour share (right graph). The
• B-: flat productivity, flat value added, significant evidence shown in Figure  2 is that
increasing prices, decreasing hourly wages, the clusters in which productivity was rising
decreasing stable hours worked and showed a decrease in their relative share and
decreasing total wages. This cluster in- that only the A++ cluster increased its relative
cluded sectors in both industry and services: weight although not in all countries.
construction, education, transportation and Notwithstanding the positive intra-industry
storage were included in this cluster in al- effect of increased productivity on employment,
most every country. It is worth noting that the increase in occupation in cluster A++ (less
all the manufacturing sectors in this cluster than 2% in relative terms) was smaller than
were low tech. The inclusion of education, that in cluster B+ (more than 10% in relative
financial and insurance activities was not terms), where productivity was stagnant (which
surprising. In most cases, because educa- may explain the above-mentioned stagnation
tion systems are part of the wider public of productivity). Moreover, the A- and A+ clus-
sphere, they were negatively affected by ters, which ranged respectively from 5% to 15%

Denmark France Germany Italy UK US Denmark France Germany Italy UK US

95 100 105

60
60

110
100
50
50

90
40
40

30
30

20
20

10
10
0

2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015

B+ A++ A+ B+ A++ A+ Total real wages, 2000=100, right scale

B− A− Total hours, 2000=100, right scale B− A− wages/value added, 2000=100, right scale

Figure 2.  Labour dislocation: clusters shares of hours worked (left) and wages (right).
Note: Authors’ elaboration from the STAN database.

164
 Are machines stealing our jobs?

of the total hours worked in 2000, showed a dislocation. To the best of our knowledge, IFR
clearly declining path, thus supporting the la- is the most comprehensive database on installed
bour dislocation literature. Moreover, cluster industrial robots per sector and country since

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

A++ showed the largest wage increase as well the 1990s. Nevertheless, data on some countries
as a stagnant relative weight (right graph), and sectors are missing: for example, data on the
which corroborates the hypotheses of the skill USA are available only since 2011, while prior to
bias (Acemoglu, 2002) and Great Decoupling that year, they are provided jointly with data on
(Brynjolfsson and McAfee, 2014) literature. Canada.
Only a few workers in each country (between To find evidence of the correlation be-
5% and 20%) benefited by the positive return tween robotisation and cluster dynamics at
from technology and most employees experi- the country–sector level, we focussed on data
enced flat or decreasing wage dynamics. on the period from 2011 to 2015, which were
In addressing the question about the future the closest to the time-series data used in
prevailing scenario, because the empirical evi- the cluster analysis. The analysis was based
dence in our study showed that total wages on two key indicators: robot intensity (the
followed an increasing trend in all countries number of robots installed in a given sector
(except Italy), we argue that the most pessim- divided by the number of employees in that
istic vision (that is, ECT) is not likely to occur sector in thousands) and the average growth
in the next few years. Although wages, except rate of robots. Because important differ-
during the crisis period, followed a growth ences in the process of robotisation affect the
path, this was only one factor. In line with the manufacturing and service sectors, we present
labour dislocation and technological bias the- the descriptive statistics in two tables: Table  3
ories, labour shares decreased in all countries is focussed on manufacturing, and Table  4 is
(only Denmark followed a different path). focussed on services. The descriptive statis-
Although aggregate output increased, workers’ tics at the country–sector level are provided in
purchasing power increased at a much lower Appendix A (Tables 1A and 2A).
rate than in the entire economy (Brynjolfsson Regarding the manufacturing sector, clus-
and McAfee, 2014; Compagnucci et al., 2018). ters A++ and A+, not surprisingly, showed the
The ongoing structural change in the in-
creased productivity in clusters A  (and to a
lesser extent in B +) has stimulated economic Table 3.  Manufacturing, clusters and robots
growth but also reduced employment (in terms
of hours worked) and has consequently de- Robot intensity in Robot average
creased the wages of workers who are expelled 2015 = robot/ growth rates
employees between
from clusters A, causing increased inequality (thousands) 2011 and 2015
and reducing labour shares.
A++ Mean 47 3.54
Median 37 3.29
Fantastic robots and where to A+ Mean 17 7.09
find them Median 10 2.43
A- Mean 8 13.41
In the previous section, we sought to identify Median 2 6.38
whether there were common dynamics among B+ Mean 3 −2.16
sectors and countries based on STAN data. In Median 3 −2.16
this section, we use data from IFR to focus on B- Mean 13 10.38
Median 6 13.13
the specific role of industrial robots in enhancing
productivity and eventually causing labour Note: Authors’ elaboration from the IFR database.

165
Gentili et al.

Table 4.  Services, clusters and robots

Robot intensity Robot average Percentage of service sectors in cluster

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

in 2015 = robot/ growth rates between
employee (thousands) 2011 and 2015 B+ A++ A+ B-

Denmark 0.002 n/a 54% 0 0 36%

France 0.002 44.58 63% 9% 18% 9%
Germany 0.015 47.14 36% 9% 18% 36%
Italy 0.003 51.87 45% 9% 0 45%
UK 0.001 32.46 54% 18% 0 27%
USA 0.002 33.34 72% 9% 0 18%

Note: Authors’ elaboration from the IFR database.

highest level of robotisation. Both clusters were (for example, the construction sector after the
characterised by the highest growth in prod- 2007 crisis), or they could have had a measure-
uctivity, although there were important differ- ment problem (for example, the outcome in
ences. As discussed in the previous section, A++ the education sector was difficult to measure).
showed increasing productivity coupled with Moreover, because productivity was given by
increasing employment and labour remuner- the ratio between output value added and em-
ation. These peculiarities, however, were shown ployees when robotisation affected traditional
in a very limited number of sectors: electrical, industries, increased productivity was limited
electronic and optical equipment in Germany; (for example, food, beverages and tobacco).
transport equipment in Germany and the UK This cluster, which included very different in-
(Table  2). The A+ cluster, faced with raising dustries, requires further investigation.
productivity, experienced increased wages in The role of robotisation in the service sector
line with increased productivity, but the work- was difficult to assess. In the IFR database, all
force decreased. Finally, A- and, to a lesser industries in the service sector are in only two
extent, B- had the lowest intensity of robots al- categories: all other non-manufacturing, and
though the latter showed the highest increase education/research. Information about the
in the number of robots. former is provided in Table  5. It was not sur-
Although causal inferences are not appro- prising that the evidence of robotisation was
priate when the panel of data is limited, these not comparable with the values observed in
results suggest that in manufacturing clusters manufacturing because services are generally
with the highest robotisation growth rates the provided by human activity. However, other
number of hours worked was decreased: A+, A- results were significant. First, there were im-
and B- showed this common trend. Nonetheless, portant differences in the relative weight of
the cluster that expanded in terms of both em- each cluster among countries. Second, the
ployment and wages was also the cluster most growth rate of robotisation in this macro-group
significantly affected by the robotisation pro- was particularly high. While it is quite normal
cess. Cluster B- experienced the highest robot to have high growth rates starting from values
growth rate and the worst productivity per- close to 0, the potential expansion of robotisa-
formance and decrease in employment. This tion in these sectors is huge. Because of the de-
result suggests that the economic activities velopment of AI, it is likely that these sectors
included in this cluster may have been nega- would also be substantially affected by auto-
tively affected by a fall in aggregate demand mation and would have the same consequences

166
 Are machines stealing our jobs?

Conclusion
Table 5.  Education and robots
The results of previous theoretical and em-
pirical investigations differ regarding the

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

Robot intensity Robot average Cluster effects of robotisation and AI on product-
in 2015 = robot/ growth rates ivity, employment and wages. Both optimists
employee between 2011
(thousands) and 2015
and pessimists, however, agree on the per-
vasiveness of the structural changes that
Denmark 0.3 0.02 B− economies now face because of robotisation
France 0.1 22.21 B− and AI. In this work, we provided empirical
Germany 0.8 −1.52 B−
evidence that both positive and negative
Italy 0.1 8.79 B−
UK 0.1 3.69 B− scenarios could result from different eco-
USA 0.1 35.93 B nomic structures with different economic
specialisations. This finding implies differ-
Note: Authors’ elaboration from IFR database.
ences in the ability and capacity to adopt
new technologies. Moreover, different out-
comes result from the peculiar and complex
experienced in the manufacturing workforce. (Myrdal, 1957) circle of causation among
Because services involved more than 60% of productivity, employment, wages and aggre-
the total workforce in each country, the poten- gate demand. This further implies that this
tial consequences for the labour market are process is time-sensitive and depends on the
dramatic. Therefore, policymakers should place current stage of development in a country
this issue high on their agenda. as well as its socio-institutional features. The
Finally, education (Table  5) was included in latter is not considered in this work.
the B- cluster in all six countries. This evidence In performing the cross-country and cross-
appears to counter the fact that modern soci- sector cluster analysis, we showed that the re-
eties are increasingly driven by the knowledge sulting country–sector pairs were synthesised
economy. Three main factors could explain this in five main clusters. The first cluster (A++) was
result. On the one hand, the measurement of the most robotised, showing that increasing
productivity in education is one of the most productivity may be coupled with increasing
complex and probably less reliable operations labour utilisation and compensation, which is
in a cross-country analysis (Bradley et al., 2010; in line with the optimistic vision of the future.
National Research Council and Sullivan, 2012). Nonetheless, this cluster, which was mainly
On the other hand, education, as previously ex- composed of high technology industries, was by
plained, is mainly publicly funded (at least in far the smallest in terms of the share of employ-
EU countries) and was negatively affected by ment. The second and third clusters (A+ and
the austerity policies following the 2007 crisis. A-) showed that while productivity increased
Third, education is strictly related to popula- with robotisation, labour utilisation decreased.
tion growth and the demand for education. The sectors in these clusters may experience
These two variables, depending on the period labour dislocation; however, in A+, the wage
analysed, could be negatively affected by the increase was large enough to compensate for
decreasing demand for education due to low the decreasing trend in employment. In A-, the
birth rates. Nonetheless, further investigations halving of labour utilisation negatively affected
of the implications of robotisation in this sector total wages. Finally, clusters B+ and B- included
may be needed. the stagnant sectors in terms of productivity.

167
Gentili et al.

Nonetheless, B-, which experienced faster ro- of the fastest economic growth in Europe’s history
botisation, showed labour dislocation. (Crafts and Toniolo, 2012).
The ‘doomsayers’ perspective’ seems far from 6
UK National Statistical Office.

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

being realised because total wages increased in 7
For instance, Italy—‘Food Products, Beverages and
all countries except Italy. Overall, the results Tobacco’, Germany—‘Food Products, Beverages and
were in line with the optimistic vision. Clusters Tobacco’, Italy—‘Transport Equipment’, etc.
A- and A+ experienced the fastest growth in 8
Data on Japan are not complete in the STAN
both robotisation and labour dislocation, which database.
affected about 15% of the total labour force. 9
According to Eurostat (2009), in 2006, Germany
This percentage may increase in the short to was ranked first in the manufacture of motor ve-
medium term because of the rapid expansion hicles, trailers and semi-trailers (NACE Division
in the robotisation of sectors in cluster B, which 34) in terms of value added (68,225 million Euro—
at present employ more than 25% of the total four times the value added of France, which ranked
labour force. second at 47.4% of EU27 value added in this sector)
Although robots will not completely replace and persons employed (840,400 employees—three
the human workforce in the short run, the issue times the French workforce, which ranked second at
37.6% of EU27 employment in this sector).
of labour dislocation must be addressed by tar-
geted policies because of its negative effects
on employment and wealth polarisation in our Acknowledgements
countries. Specifically, because robotisation This work was supported by the Università Politecnica
seems to be growing faster than the capacity delle Marche and the Università di Macerata. We thank
of workers to acquire new skills, policies based the anonymous reviewers for their several constructive
on sustained income in conjunction with adult suggestions, which improved our paper.
learning will be increasingly necessary for the
near future.
References
Acemoglu,  D. (2002) Technical change, inequality,
Endnotes and the labour market, The Journal of Economic
Literature, 40: 7–72.
1
Dosi and Virgillito (2019) discussing Freeman Acemoglu, D. (2015) Localised and biased technolo-
(1992) and Freeman and Soete (1994). gies: Atkinson and Stiglitz’s new view, induced in-
novations, and directed technological change, The
2
According to Bárány and Siegel (2018) and the Economic Journal, 125: 443–463.
OECD (2018b), manufacturing, among the eco- Acemoglu,  D. and Restrepo,  P. (2018) The race be-
nomic sectors, was most affected by automation in tween man and machine: implications of tech-
the recent past and is likely to be most affected in the nology for growth, factor shares, and employment,
near future (Manyika, 2017). American Economic Review, 108: 1488–1542.
3
The multiplier effect is the effect according to Arntz, M., Gregory, T. and Zierahn, U. (2016) The risk
of automation for jobs in OECD countries: a com-
which each new high-tech job in the USA creates
parative analysis, OECD Social, Employment and
five additional jobs in the service economy. Migration Working Papers, 189: 1–34.
4
Therefore, reverse causality in the relation cannot be Autor, D. and Dorn, D. (2013) The growth of low skill
excluded because robotisation can be considered the service jobs and the polarization of the US labor
reaction of firms to increasing salaries and labour util- market, American Economic Review, 103: 1553–1597.
isation. This hypothesis, however, seems less plausible Autor, D., Levy, F. and Murnane, R. (2003) The skill
given the ongoing reduction of labour shares. content of recent technological change: an empir-
ical exploration, Quarterly Journal of Economics,
5
The French economist Jean Fourastié called the 118: 1279–1333.
years between the late 1940s and the early 1970s Autor,  D. and Salomons,  A. (2018) Is automation
‘les trente glorieuses’, which stand out as the period labor share-displacing? Productivity growth,

168
 Are machines stealing our jobs?

employment, and the labor share, Brookings Centre for Economics and Business Research
Papers on Economic Activity, 2018: 1–87. (CEBR) (2017) The Impact of Automation:
Bárány,  Z.  L. and Siegel,  C. (2018) Job polariza- A Report for Redwood. Available online at: https://
tion and structural change, American Economic cebr.com/wp-content/uploads/2017/03/Impact_

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

Journal: Macroeconomics, 10: 57–89. of_automation_report_23_01_2017_FINAL.pdf.
Baumol, W. J. (1967) Macroeconomics of unbalanced Chandrasekhar,  P. (2018) ‘Wageless Growth’ Not
growth: the anatomy of urban crisis, American ‘Jobless Growth’: The New Conundrum. Available
Economic Review, 57: 415–426. online at: https://www.networkideas.org/news-
Beck, P., Hardie, M., Jones, N. and Loakes, A. (2017) analysis/2018/12/wageless-growth-not-jobless-
The Feasibility of Measuring the Sharing Economy: growth-the-new-conundrum/.
November 2017 Progress Update. United Cirillo,  V., Pianta,  M. and Nascia,  L. (2015) The
Kingdom Office for National Statistics. Available Dynamics of Skills: Technology and Business
online at https://www.ons.gov.uk/economy/ Cycles. LEM Papers Series 2015/30. Laboratory
economicoutputandproductivity/output/articles/ of Economics and Management (LEM).
thefeasibilityofmeasuringthesharingeconomy/ Available online at: https://ideas.repec.org/p/ssa/
november2017progressupdate. lemwps/2015–30.html.
Bennion,  E.  G. (1943) Unemployment in the the- Compagnucci,  F., Gentili,  F., Valentini,  E. and
ories of Schumpeter and Keynes, The American Gallegati, M. (2018) Have jobs and wages stopped
Economic Review, 33: 336–347. rising? Productivity and structural change in
Bessen,  J.  E. (2019) Automation and Jobs: When advanced countries. In Structural Change and
Technology Boosts Employment. Boston Economic Dynamics. Available online at: https://
University School of Law, Law and Economics www.sciencedirect.com/science/article/abs/pii/
Research Paper No. 17-09. Available online at: S0954349X17301832.
https://ssrn.com/abstract=2935003 or http://dx.doi. Compagnucci,  F., Gentili,  A., Valentini,  E. and
org/10.2139/ssrn.2935003. Gallegati, M. (2019) Robotization and labour dis-
Bessen, J. E. (2018) AI and Jobs: The Role of Demand. location in the manufacturing sectors of OECD
Boston University School of Law, Law and countries: a panel VAR approach, Applied
Economics Research Paper No. 17–46. Available Economics, 51: 6127–6138.
online at: https://ssrn.com/abstract=3078715 or Crafts,  N. and Toniolo,  G. (2012) ‘Les trente
http://dx.doi.org/10.2139/ssrn.3078715. glorieuses’: from the Marshall plan to the oil
Boston Consulting Group (2015) The Robotics crisis. In: D.  Stone (ed.), The Oxford Handbook
Revolution. The Next Great Leap in Manufacturing. of Postwar European History. Oxford: Oxford
Available online at: https://circabc.europa.eu/sd/a/ University Press.
b3067f4e-ea5e-4864-9693-0645e5cbc053/BCG_The_ Dauth,  W., Findeisen,  S., Suedekum,  J. and
Robotics_Revolution_Sep_2015_tcm80-197133.pdf. Woessner, N. (2018) Adjusting to Robots: Worker
Bradley,  S., Johnes,  J. and Little,  A. (2010) Level Evidence. Institute Working Paper 13.
Measurement and determinants of efficiency and Deutschen Instituts für Wirtschaftsforschung –
productivity in the further education sector in DIW Berlin (Institut für Konjunkturforschung),
England, Bulletin of Economic Research, 62: 1–30. Opportunity and Inclusive Growth Institute.
Breazeal,  C., Takanishi,  A. and Kobayashi,  T. (2008) Decker, M., Fischer, M. and Ott, I. (2017) Service ro-
Social robots that interact with people. In B. Siciliano botics and human labor: a first technology assess-
and O.  Khatib (eds.), Springer Handbook of ment of substitution and cooperation, Robotics
Robotics. Berlin and Heidelberg: Springer. and Autonomous Systems, 87: 348–354.
Brynjolfsson, E. and McAfee, A. (2014) The Second Delli  Gatti,  D., Gallegati,  M., Greenwald,  B.  C.,
Machine Age: Work, Progress, and Prosperity in Russo,  A. and Stiglitz,  J.  E. (2012) Mobility con-
a Time of Brilliant Technologies. New York, NY: straints, productivity trends, and extended crises,
WW Norton & Company. Journal of Economic Behavior & Organization,
Brynjolfsson,  E., Rock,  D. and Syverson,  C. (2017) 83: 375–393.
Artificial Intelligence and the Modern Productivity Dosi,  G. and Virgillito,  M.  E. (2019) Whither the
Paradox: A Clash of Expectations and Statistics. No. Evolution of the Contemporary Social Fabric?
w24001. National Bureau of Economic Research. New Technologies and Old Socio-economic Trends.
Camagni,  R. (2017) Spatial diffusion of pervasive Global Labor Organization Discussion Paper No.
process innovation. In Seminal Studies in Regional 316. Available online at: https://www.econstor.eu/
and Urban Economics, pp. 49–63. Cham: Springer. handle/10419/191764.

169
Gentili et al.

European Commission (EC) (2016) Analysis of org/active/publications/discussion_papers/

the Impact of Robotic Systems on Employment dp.php?dpno=10477.
in the European Union—2012 Data Update. Graetz,  G. and Michaels,  G. (2017) Is modern
Luxembourg: Publications Office of the European technology responsible for jobless recoveries?,

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

Union. American Economic Review, 107: 168–173.
European Commission (EC) (2019) The Future Green,  F. and Henseke,  G. (2016) Should govern-
of Work? Work of the Future! On How ments of OECD countries worry about graduate
Artificial Intelligence, Robotics and Automation underemployment?, Oxford Review of Economic
Are Transforming Jobs and the Economy in Policy, 32: 514–537.
Europe. doi:10.2872/913422. Available on- Gries,  T. and Naudé,  W. (2018) Artificial intelligence,
line at: https://ec.europa.eu/epsc/publications/ jobs, inequality and productivity: Does aggregate de-
other-publications/future-work-work-future_en. mand matter?, IZA Discussion Paper, 12005: 1–39.
Eurostat (2009) European Business: Facts and Harari,  Y.  N. (2014) Sapiens: A  Brief History of
Figures. Luxembourg: Office for Official Humankind. London: Harvill Secker, Penguin
Publications of the European Communities. Random House.
doi:10.2785/23246 Huang, M. H. and Rust, R. T. (2018) Artificial intel-
Felten,  E.  W., Raj,  M. and Seamans,  R. (2008) ligence in service, Journal of Service Research, 21:
A method to link advances in artificial intelli- 155–172.
gence to occupational abilities, AEA Papers and International Federation of Robotics (2018) World
Proceedings, 108: 1–4. Robotics Industrial Robots Database. Available
Frank, M. R., Autor, D., Bessen, J. E., Brynjolfsson, online at: www.ifr.org [Accessed October 2018].
E., Cebrian, M., Deming, D. J., Feldman, M., Groh, Kontsevaya,  S. (2017) Analysis and control of gov-
M., Lobo, J., Moro, E., Wang, D., Youn, H. and ernment subsidies for investments in agricul-
Rahwan, P. (2019) Toward understanding the im- ture of the Russian Federation, Acta Scientiarum
pact of artificial intelligence on labor, Proceedings Polonorum Oeconomia, 16:  25–33.
of the National Academy of Sciences of the United Korinek, A. and Stiglitz, J. E. (2017) Artificial intel-
States of America, 116: 6531–6539. ligence, worker replacing technological progress
Freeman, C. (1992) The Economics of Hope: Essays and income distribution, NBER Working Paper,
on Technical Change and Economic Growth. 24174.
London: Pinter. Kromann, L., Skaksen, J. R. and Sørensen, A. (2011)
Freeman,  C. and Soete,  L. (1994) Work for All or Automation, Labor Productivity and Employment
Mass Unemployment? Computerised Technical – A Cross country Comparison. CEBR,
Change into the 21st Century. London and New Copenhagen Business School. Available online at:
York: Pinter Publisher. https://ssrn.com/abstract=3096432 or http://dx.doi.
Frey, C. B. and Osborne, M. A. (2017) The future of org/10.2139/ssrn.3096432.
employment: how susceptible are jobs to compu- Leontief,  W. (1952) Machines and man, Scientific
terisation? Technological Forecasting and Social American, 187: 150–164.
Change, 114: 254–280. Mandfield,  E. (1989) The diffusion of industrial
Giaoutzi, M. and Stratigea, A. (1991) The impact of robots in Japan and the United States, Research
new information technologies on spatial inequal- Policy, 18: 183–192.
ities. In Regional Science, pp. 191–210. Berlin and Makles, A. (2012) Stata tip 110: how to get the op-
Heidelberg: Springer. timal kmeans cluster solution, The Stata Journal,
Goldin, I., Koutroumpis, P., Lafond, F., Rochowicz, N. 12: 347–351.
and Winkler, J. (2018) Why Is Productivity Slowing Makridakis,  S. (2017) The forthcoming artificial in-
Down? Working Paper. Oxford: Oxford Martin telligence (AI) revolution: its impact on society
School, University of Oxford. and firms, Futures, 90: 46–60.
González,  J.  C., Pulido,  J.  C. and Fernández,  F. Manyika, J. (2017) A Future That Works: AI
(2017) A three-layer planning architecture for Automation Employment and Productivity.
the autonomous control of rehabilitation ther- Technical Report. McKinsey Global Institute
apies based on social robots, Cognitive Systems Research. Available online at: https://www.
Research, 43: 232–249. mckinsey.com/~/media/mckinsey/featured%20
Graetz,  G. and Michaels,  G. (2015) ‘Robots at insights/Digital%20Disruption/Harnessing%20
Work’. London: Centre for Economic Policy automation%20for%20a%20future%20that%20
Research. Available online at: https://cepr. works/MGI-A-future-that-works-Full-report.ashx.

170
 Are machines stealing our jobs?

Meliciani,  V. (2001) Technology, Trade and Growth OECD. Available online at: www.oecd.org/going-
in OECD Countries: Does Specialisation Matter? digital/productivity-growth-in-the-digital-age.pdf.
Routledge Studies in the Modern World Economy. Pew Research Center (2017) Automation in Everyday
London: Taylor & Francis. Life. Available online at: http://www.pewinternet.

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

Mokyr,  J. (2002) The Gifts of Athena: Historical o r g / w p - c o n t e n t / u p l o a d s / s i t e s / 9 / 2 0 17 / 1 0 /
Origins of the Knowledge Economy. Princeton, PI_2017.10.04_Automation_FINAL.pdf.
NJ: Princeton University Press. Piketty, T. and Zucman, G. (2014) Capital is back:
Mokyr, J., Vickers, C. and Ziebarth, N. L. (2015) The wealth income ratios in rich countries 1700–
history of technological anxiety and the future 2010, The Quarterly Journal of Economics, 129:
of economic growth: is this time different? The 1255–1310.
Journal of Economic Perspectives, 29: 31–50. Piva,  M. and Vivarelli,  M. (2018) Technological
Moretti, E. (2012) The New Geography of Jobs. New change and employment: is Europe ready for
York: Houghton Mifflin Harcourt. the challenge?, Eurasian Business Review, 8:
Myrdal,  G. (1957) Economic Theory and Under­ 13–32.
developed Regions. New York: Harper and Row. Repkine, A. (2012) How similar are the East Asian
Nasri, A. and Zhang, L. (2018) A multi-dimensional Economies? A  cluster analysis perspective on
multi-level approach to measuring the spatial economic cooperation in the region, Journal of
structure of U.S.  metropolitan areas, Journal of International and Area Studies, 19: 27–44.
Transport and Land Use, 11: 49–65. Saint-Paul,  G. (2017). Robots: vers la fin du travail?
National Research Council and Sullivan, T. A. (2012) Available online at: https://halshs.archives-ouvertes.
Improving Measurement of Productivity in Higher fr/halshs-01513006/document.
Education. Washington, DC: National Academies Sujarwoto, S. and Tampubolon, G. (2016) Spatial in-
Press. equality and the Internet divide in Indonesia 2010–
Hägele, M., Nilsson, K., Pires, J. N. and Bischoff, R. 2012, Telecommunications Policy, 40: 602–616.
(2016) Industrial robotics. In Springer Handbook Theis, W. and Weihs, C. (2000) Clustering techniques
of Robotics, pp. 1385–1422. Cham: Springer. for the detection of business cycles. In R. Decker
OECD (2011) Recent Developments in the and W. Gaul (eds.), Classification and Information
Automobile Industry. OECD Economics Processing at the Turn of the Millennium, pp. 127–
Department Policy Notes, No. 7. Available on- 134. Berlin and Heidelberg: Springer.
line at: https://www.oecd.org/economy/out- United Nations (2017) The Slowdown in
look/48333701.pdf. Productivity Growth: A  View from International
OECD (2012) OECD Employment Outlook 2012. Trade. Development Issues No. 11. Available on-
Paris: OECD Publishing. line at:  https://www.un.org/development/desa/
OECD (2018a) OECD Employment Outlook 2018. dpad/wpcontent/uploads/sites/45/publication/
Paris: OECD Publishing. dsp_policy_11.pdf.
OECD (2018b) Job Creation and Local Economic Vivarelli,  M. (2014) Innovation, employment and
Development 2018: Preparing for the Future of skills in advanced and developing countries: a
Work. Paris: OECD Publishing. survey of economic literature, Journal of Economic
OECD (2018c) OECD Science, Technology and Issues, 48: 123–154.
Innovation Outlook 2018:Adapting to Technological Vivarelli, M. and Pianta, M. (2000) The Employment
and Societal Disruption. Paris: OECD Publishing. Impact of Innovation: Evidence and Policy.
OECD (2019) Productivity Growth in the Digital Routledge Studies in the Modern World Economy.
Age, OECD Going Digital Policy Note. Paris: London: Routledge.

171
Gentili et al.

Appendix A

Table 1A.  Robot intensity and robot average growth rate: Denmark, France and Germany

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

Robot Robot average
intensity 2015 growth rates
(2011–2015)

Denmark Construction 0.21 +24.71%

Basic metals and fabricated metal products; machinery and equipment 41.01 −1.22%
Electrical, electronic and optical equipment 15.52 −1.13%
Food products, beverages and tobacco 17.33 +10.25%
Furniture; other manufacturing; repair and installation of machinery 2.04 −1.21%
and equipment
Plastic and chemical products, glass, ceramics, stone, mineral products 19.63 +10.11%
Textiles, wearing apparel, leather and related products 27.33 +6.13%
Transport equipment 35.40 +9.79%
Wood and paper products, and printing 14.34 −6.37%
Education 0.31 +0.02%
All other services 0.002 .
France Construction 0.14 +23.10%
Basic metals and fabricated metal products; machinery and equipment 10,27 +0.73%
Electrical, electronic and optical equipment 5.52 −0.40%
Food products. beverages and tobacco 5.15 +7.25%
Furniture; other manufacturing; repair and installation of machinery 1.84 +9.37%
and equipment
Plastic and chemical products, glass, ceramics, stone, mineral products 10.75 −0.06%
Textiles, wearing apparel, leather and related products 0.47 −1.68%
Transport equipment 76.55 −6.70%
Wood and paper products, and printing 3.34 +9.42%
Education 0.13 +22.21%
All other services 0.002 +44.58%
Germany Construction 0.09 +10.32%
Basic metals and fabricated metal products; machinery and equipment 20.26 +5.03%
Electrical, electronic and optical equipment 10.17 +2.27%
Food products, beverages and tobacco 7.69 +3.50%
Furniture; other manufacturing; repair and installation of machinery 3.65 −10.90%
and equipment
Plastic and chemical products, glass, ceramics, stone, mineral products 18.91 +2.43%
Textiles, wearing apparel, leather and related products 1.81 +7.13%
Transport equipment 93.52 +3.29%
Wood and paper products, and printing 3.19 −10.82%
Education 0.775 1.52%
All other services 0.015 +47.19%

Note: Authors’ elaboration from IFR database.

172
 Are machines stealing our jobs?

Table 2A.  Robot intensity and robot average growth rate: Italy, UK and USA

Downloaded from https://academic.oup.com/cjres/article/13/1/153/5731513 by University of Tasmania Library user on 27 May 2021

Robot Robot average
intensity 2015 growth rates
(2011–2015)

Italy Construction 0.22 +13.12%

Basic metals and fabricated metal products; machinery and equipment 25.83 +2.56%
Electrical, electronic and optical equipment 6.43 +1.57%
Food products, beverages and tobacco 14.80 +20.86%
Furniture; other manufacturing; repair and installation of machinery and 3.11 +2.84%
equipment
Plastic and chemical products, glass, ceramics, stone, mineral products 21.23 −8.74%
Textiles, wearing apparel, leather and related products 0.65 −2.07%
Transport equipment 61.89 −5.26%
Wood and paper products, and printing 4.41 +8.09%
Education 0.10 +8.78%
All other services 0.003 +51.87%
UK Construction 0.07 −11.57%
Basic metals and fabricated metal products; machinery and equipment 4.10 +11.08%
Electrical, electronic and optical equipment 2.34 +0.71%
Food products, beverages and tobacco 2.95 +8.42%
Furniture; other manufacturing; repair and installation of machinery and 1.14 +6.84%
equipment
Plastic and chemical products, glass, ceramics, stone, mineral products 6.95 +6.38%
Textiles, wearing apparel, leather and related products 0.16 +1.79%
Transport equipment 36.71 +5.06%
Wood and paper products, and printing 0.41 −7.68%
Education 0.08 +3.69%
All other services 0.001 +32.46%
USA Construction 0.19 +26.70%
Basic metals and fabricated metal products; machinery and equipment 10.88 +15.94%
Electrical, electronic and optical equipment 25.73 +17.37%
Food products, beverages and tobacco 5.63 +14.15%
Furniture; other manufacturing; repair and installation of machinery and 7.09 +64.87%
equipment
Plastic and chemical products, glass, ceramics, stone, mineral products 9.61 +14.20%
Textiles, wearing apparel, leather and related products 0.28 +60.10%
Transport equipment 69.23 +15.32%
Wood and paper products, and printing 0.35 +126.21%
Education 0.05 +35.93%
All other services 0.002 +33.34%

Note: Authors’ elaboration from IFR database.

173