by Ajay Agrawal
ing, welding, painting, palletizing, assembly, material handling, and quality
control (Ayres and Miller 1983; Groover et al. 1986; Graetz and Michaels
2015; Acemoglu and Restrepo 2017).
Examples of automation are not confi ned to industry and agriculture.
Computer software has already automated a number of tasks performed by
white- collar workers in retail, wholesale, and business services. Software and
AI- powered technologies can now retrieve information, coordinate logis-
tics, handle inventories, prepare taxes, provide fi nancial services, translate
complex documents, write business reports, prepare legal briefs, and diag-
nose diseases. These technologies are set to become much better at these
and other tasks during the next years (e.g., Brynjolfsson and McAfee 2014;
Ford 2016).
As these examples illustrate, automation involves the substitution of
machines for labor and leads to the displacement of workers from the tasks
that are being automated. This displacement eff ect is not present—or pres-
ent only incidentally—in most approaches to production functions and
labor demand used in macroeconomics and labor economics. The canoni-
cal approach posits that production in the aggregate (or in a sector for that
matter) can be represented by a function of the form F( AL, BK), where L
denotes labor and K is capital. Technology is assumed to take a “factor-
augmenting” form, meaning that it multiplies these two factors of produc-
tion as the parameters A and B do in this production function.
It might appear natural to model automation as an increase in B, that is,
as capital- augmenting technological change. However, this type of techno-
logical change does not cause any displacement and always increases labor
demand and wages (see Acemoglu and Restrepo 2016). Moreover, as our
examples above illustrate, automation is not mainly about the development
of more productive vintages of existing machines, but involves the intro-
202 Daron Acemoglu and Pascual Restrepo
duction of new machinery to perform tasks that were previously the domain
of human labor.
Labor- augmenting technological change, corresponding to an increase
in A, does create a type of displacement if the elasticity of substitution
between capital and labor is small. But in general, this type of technologi-
cal change also expands labor demand, especially if capital adjusts over the
long run (see Acemoglu and Restrepo 2016). Moreover, our examples make
it clear that automation does not directly augment labor; on the contrary,
it transforms the production process in a way that allows more tasks to be
performed by machines.
8.2.1 Tasks, Technologies, and Displacement
We propose, instead, a task- based approach, where the central unit of
production is a task as in the textile example discussed above.3 Some tasks
have to be produced by labor, while other tasks can be produced either by
labor or by capital. Also, labor and capital have comparative advantages in
diff erent tasks, meaning that the relative productivity of labor varies across
tasks. Our framework conceptualizes automation (or automation at the
extensive margin) as an expansion in the set of tasks that can be produced
with capital. If capital is suffi
ciently cheap or suffi
ciently productive at the
margin, then automation will lead to the substitution of capital for labor
in these tasks. This substitution results in a displacement of workers from
the tasks that are being automated, creating the aforementioned displace-
ment eff ect.
The displacement eff ect could cause a decline in the demand for labor and
the equilibrium wage rate. The possibility that technological improvements
that increase productivity can actually reduce the wage of all workers is an
important point to emphasize because it is often downplayed or ignored.
With an elastic labor supply (or quasi- labor supply refl ecting some labor
market imperfections), a reduction in the demand for labor also leads to
lower employment. In contrast to the standard approach based on factor-
augmenting technological changes, a task- based approach immediately
opens the way to productivity- enhancing technological developments that
simultaneously reduce wages and employment.
8.2.2 Countervailing
Eff ects
The presence of the displacement eff ect does not mean that automation
will always reduce labor demand. In fact, throughout history, there are
several periods where automation was accompanied by an expansion of
3. See Autor, Leavy, and Murnane (2003) and Acemoglu and Autor (2011). Diff erent from
these papers that develop a task- based approach focusing on inequality implications of technological change, we are concerned here with automation and the process of capital- replacing tasks previously performed by labor and their implications for wages and employment.
Artifi cial Intelligence, Automation, and Work 203
labor demand and even higher wages. There are a number of reasons why
automation could increase labor demand.
1. The Productivity Eff ect. By reducing the cost of producing a subset of
tasks, automation raises the demand for labor in nonautomated tasks (Autor
2015; Acemoglu and Restrepo 2016). In particular, automation leads to the
substitution of capital for labor because at the margin, capital performs
certain tasks more cheaply than labor used to. This reduces the prices of the
goods and services whose production processes are being automated, mak-
ing households eff ectively richer, and increasing the demand for all goods
and services.
The productivity eff ect could manifest itself in two complementary ways.
First, labor demand might expand in the same sectors that are undergoing
automation.4 A telling example of this process comes from the eff ects of the
introduction of automated teller machines (ATMs) on the employment of
bank tellers. Bessen (2016) documents that concurrent with the rapid spread
of ATMs—a clear example of automating technology that enabled these
new machines to perform tasks that were previously performed more expen-
sively by labor—there was an expansion in the employment of bank tellers.
Bessen suggests that this is because ATMs reduced the costs of banking and
encouraged banks to open more branches, raising the demand for bank tell-
ers who then specialized in tasks that ATMs did not automate.
Another interesting example of this process is provided by the dynam-
ics of labor demand in spinning and weaving during the British Industrial
Revolution as recounted by Mantoux (1928). Automation in weaving (most
notably, John Kay’s fl y shuttle) made this task cheaper and increased the
price of yarn and the demand for the complementary task of spinning.
Later automation in spinning reversed this trend and increased the demand
for weavers. In the words of John Wyatt, one of the inventors of the spin-
ning machine, installing spinning machines would cause clothiers to “then
want more hands in every other branch of the trade, viz. weavers, shearmen,
scourers, combers, etc.” (quoted in Mantoux 1928). T
his is also probably
the reason why the introduction of Eli Whitney’s cotton gin in 1793, which
automated the labor- intensive process of separating the cotton lint from
the seeds, appears to have led to greater demand for slave labor in southern
plantations (Rasmussen 1982).
The productivity eff ect also leads to higher real incomes and thus to greater
demand for all products, including those not experiencing automation. The
greater demand for labor from other industries might then counteract the
negative displacement eff ect of automation. The clearest historical example
of this comes from the adjustment of the US and many European economies
4. This requires that the demand for the products of these sectors is elastic. Acemoglu and Restrepo (2017) refer to this channel as the price- productivity eff ect because it works by reducing the relative price of products that are being automated and restructuring production toward these sectors.
204 Daron Acemoglu and Pascual Restrepo
to the mechanization of agriculture. By reducing food prices, mechanization
enriched consumers who then demanded more nonagricultural goods (Her-
rendorf, Rogerson, and Valentinyi 2013), and created employment oppor-
tunities for many of the workers dislocated by the mechanization process
in the fi rst place.5
This discussion also implies that, in contrast to the popular emphasis on
the negative labor market consequences of “brilliant” and highly productive
new technologies set to replace labor (e.g., Brynjolfsson and McAfee 2014;
Ford 2016), the real danger for labor may come not from highly productive
but from “so- so” automation technologies that are just productive enough
to be adopted and cause displacement, but not suffi
ciently productive to
bring about powerful productivity eff ects.
2. Capital Accumulation. As our framework in the next section clarifi es,
automation corresponds to an increase in the capital intensity of produc-
tion. The high demand for capital triggers further accumulation of capital
(e.g., by increasing the rental rate of capital). Capital accumulation then
raises the demand for labor. This may have been an important channel of
adjustment of the British economy during the Industrial Revolution and of
the American economy in the fi rst half of the twentieth century in the face
of mechanization of agriculture, for in both cases there was rapid capital
accumulation (Allen 2009; Olmstead and Rhode 2001).
As we discuss in the next section, under some (albeit restrictive) assump-
tions often adopted in neoclassical models of economic growth, capital accu-
mulation can be suffi
ciently powerful that automation will always increase
wages in the long run (see Acemoglu and Restrepo 2016), though the more
robust prediction is that it will act as a countervailing eff ect.
3. Deepening of Automation. The displacement eff ect is created by auto-
mation at the extensive margin—meaning the expansion of the set of tasks
that can be produced by capital. But what happens if technological improve-
ments increase the productivity of capital in tasks that have already been
automated? This will clearly not create additional displacement because
labor was already replaced by capital in those tasks. But it will generate the
same productivity eff ects we have already pointed out above. These pro-
ductivity eff ects then raise labor demand. We refer to this facet of advances
in automation technology as the deepening of automation (or as automa-
tion at the intensive margin because it is intensifying the productive use of
machines).
A clear illustration of the role of deepening automation comes from the
introduction of new vintages of machinery replacing older vintages used in
already automated tasks. For instance, in US agriculture the replacement of
5. Acemoglu and Restrepo (2017) refer to it as a “scale eff ect” because in their setting it acted in a homothetic manner, scaling up demand from all sectors, though in general it could take a nonhomothetic form.
Artifi cial Intelligence, Automation, and Work 205
horse- powered reapers and harvesters by diesel tractors increased produc-
tivity, presumably with limited additional substitution of workers in agri-
cultural tasks.6 In line with our account of the potential role of deepening
automation, agricultural productivity and wages increased rapidly starting
in the 1930s, a period that coincided with the replacement of horses by trac-
tors (Olmstead and Rhode 2001; Manuelli and Seshadri 2014).
Another example comes from the vast improvements in the effi
ciency of
numerically controlled machines used for metal cutting and processing (such
as mills and lathes), as the early vintages controlled by punched cards were
replaced by computerized models during the 1970s. The new computer-
ized machines were used in the same tasks as the previous vintages, and
so the additional displacement eff ects were probably minor. As a result,
the transition to CNC (computer numerical control) machines increased
the productivity of machinists, operators, and other workers in the industry
(Groover 1983).
The three countervailing forces we have listed here are central for under-
standing why the implications of automation are much richer than the direct
displacement eff ects might at fi rst suggest, and why automation need not
be an unadulterated negative force against the labor market fortunes of
workers. Nevertheless, there is one aspect of the displacement eff ect that is
unlikely to be undone by any of these four countervailing forces: as we show
in the next section, automation necessarily makes the production process
more capital intensive, reducing the share of labor in national income. Intui-
tively, this is because it entails the substitution of capital for tasks previously
performed by labor, thus squeezing labor into a narrower set of tasks.
If, as we have suggested, automation has been ongoing for centuries, with
or without powerful countervailing forces of the form listed here, we should
have seen a “nonbalanced” growth process with the share of labor in national
income declining steadily since the beginning of the Industrial Revolution.
That clearly has not been the case (see, e.g., Kuznets 1966; Acemoglu 2009).
This suggests that there have been other powerful forces making production
more labor intensive and balancing the eff ects of automation. This is what
we suggest in the next subsection.
8.2.3 New
Tasks
As discussed in the introduction, periods of intensive automation have
often coincided with the emergence of new jobs, activities, industries,
and tasks. In nineteenth- century Britain, for example, there was a rapid
expansion of new industries and jobs ranging from engineers, machinists,
repairmen, conductors, back- offi
ce workers, and managers involved with
6. Nevertheless, the move from horse power to tractors contributed to a decline in agricultural employment via a diff erent channel: tractors increased agricultural productivity, and because of inelastic demand, expenditure on agricultural products declined (Rasmussen 1982).
206 Daron Acemo
glu and Pascual Restrepo
the introduction and operation of new technologies (e.g., Landes 1969;
Chandler 1977; and Mokyr 1990). In early twentieth- century America, the
mechanization of agriculture coincided with a large increase in employ-
ment in new industry and factory jobs (Kuznets 1966) among others in the
burgeoning industries of farm equipment (Olmstead and Rhode 2001) and
cotton milling (Rasmussen 1982). This is not just a historical phenomenon.
As documented in Acemoglu and Restrepo (2016), from 1980 to 2010 the
introduction and expansion of new tasks and job titles explains about half
of US employment growth.
Our task- based framework highlights that the creation of new labor-
intensive tasks (tasks in which labor has a comparative advantage relative
to capital) may be the most powerful force balancing the growth process in
the face of rapid automation. Without the demand for workers from new
factory jobs, engineering, supervisory tasks, accounting, and managerial
occupations in the second half of the nineteenth and much of the twenti-
eth centuries, it would have been impossible to employ millions of workers
exiting the agricultural sector and automated labor- intensive tasks.
In the same way that automation has a displacement eff ect, we can think
of the creation of new tasks as engendering a reinstatement eff ect. In this
way, the creation of new tasks has the opposite eff ect of automation. It
always generates additional labor demand, which increases the share of
labor in national income. Consequently, one powerful way in which tech-
nological progress could be associated with a balanced growth path is via
the balancing of the impacts of automation by the creation of new tasks.
The creation of new tasks need not be an exogenous, autonomous process
unrelated to automation, AI, and robotics for at least two reasons:
1. As emphasized in Acemoglu and Restrepo (2016), rapid automation
may endogenously generate incentives for fi rms to introduce new labor-
intensive tasks. Automation running ahead of the creation of new tasks
reduces the labor share and possibly wages, making further automation less
profi table and new tasks generating employment opportunities for labor
more profi table for fi rms. Acemoglu and Restrepo (2016) show that this