by Ajay Agrawal
subsection.
The second one, discussed in this subsection, is that concurrent with the
rapid introduction of new automation technologies, we may be experiencing
a slowdown in the creation of new tasks and investments in other technolo-
gies that benefi t labor.
This explanation comes in two fl avors. First, we may be running out of
good ideas to create new jobs, sectors, and products capable of expanding
the demand for labor (e.g., Gordon 2016; Bloom et al. 2017), even if auto-
mation continues at a healthy or accelerating pace. Alternatively, the rapid
introduction of new automation technologies may redirect resources that
were devoted to other technological advances, in particular, the creation of
new tasks (see Acemoglu and Restrepo 2016). To the extent that the recent
enthusiasm—or even “frenzy”—about deep learning and some aspects of
AI can be viewed as such a redirection, our framework pinpoints a potential
powerful mechanism for slower productivity growth in the face of rapid
automation.
Both explanations hinge on the redirection of research activity from the
224 Daron Acemoglu and Pascual Restrepo
creation of new tasks to automation—in the fi rst case exogenously and in
the second for endogenous reasons. Recall from our analysis so far that the
productivity gains from new tasks in our baseline framework are given by
d ln( Y / L)
R
W
= ln
ln
> 0,
dN
( N
1)
( N )
M
L
while productivity gains from automation are
d ln( Y / L)
W
R
= ln
ln
> 0.
dI
( I )
( I )
L
M
If the former expression is greater than the latter, then the redirection of
research eff ort from the creation of new tasks toward automation, or a lower
research effi
ciency in creating new tasks, will lead to a slowdown of produc-
tivity growth, even if advances in automation are accelerating and being
adopted enthusiastically. This conclusion is strengthened if additional eff ort
devoted to automation at the expense of the creation of new tasks runs into
diminishing returns.
8.5.3 Excessive
Automation
In this subsection, we highlight the third reason for why there may be
modest productivity growth: socially excessive automation (see Acemoglu
and Restrepo 2016, 2018a).
To illustrate why our framework can generate excessive automation, we
modify the assumption that the supply of capital, K, is given, and instead
suppose that machines used in automation are produced—as intermediate
goods—using the fi nal good at a fi xed cost R. Moreover, suppose that be-
cause of subsidies to capital, accelerated depreciation allowances, tax credit
for debt- fi nanced investment or simply because of the tax cost of employing
workers, capital receives a marginal subsidy of τ > 0.
Given this subsidy, the rental rate for machines is R(1 – τ), and assump-
tion (A1) now becomes
( N )
( I )
L
>
W
> L
.
( N
1)
R(1
)
( I )
M
M
Let us now compute GDP as value added, subtracting the cost of produc-
ing machines. This gives us
GDP = Y
RK.
Suppose next that there is an increase in automation. Then we have
d GDP
dK
dK
= dY + R(1
)
R
,
dI
dI
dI
dI
K
which simplifi es to
Artifi cial Intelligence, Automation, and Work 225
d GDP
R 1
(
)
=
W
dK
ln
R
.
dI
( I )
ln
( I )
dI
L
M
Excessive automation<0
Productivity effect>0
The fi rst term is positive and captures the productivity increase generated by
automation. However, when τ > 0—so that the real cost of using capital is
distorted—we have an additional negative eff ect originating from excessive
automation.13 At the root of this negative eff ect is the fact that subsidies
induce fi rms to substitute capital for labor even when this is not socially
cost- saving (though it is privately benefi cial because of the subsidy).
This conclusion is further strengthened when there are also labor market
frictions as pointed out in section 8.2. To illustrate this point in the simplest
possible fashion, let us assume that there is a threshold J ∈( I, N) such that, when performing the tasks in [ I, J ], workers earn rents > 0 proportional to their wage in other tasks. In particular, workers are paid a wage W to
produce tasks in [ J, N ], and a wage W(1 + ) to produce tasks in ( I, J).14 Let L denote the total amount of labor allocated to the tasks in ( I, J ), and note A
that these are the workers that will be displaced by automation, that is, by a
small increase in I. Given this additional distortion, assumption (A1) now
becomes
( N )
( I )
L
>
W
> 1
L
.
( N
1)
R(1
)
1 +
( I )
M
M
The demand for labor in tasks where workers earn rents is now
L =
Y
( J
I ).
A
W (1 + )
The demand for labor in tasks where workers do not earn rents is
L
L = Y ( N
J ).
A
W
Dividing these two expressions, we obtain the equilibrium condition for L ,
A
L
J
I
A
= 1
,
L
L
1 +
N
J
A
13. We show in the appendix that K = ( Y/ R)( I – N + 1), which implies that K increases in I.
14. The assumption that there are rents only in a subset of tasks is adopted for simplicity. The same results apply (a) when there are two sectors and one of the sectors has higher rents/ wages for workers and enables automation and (b) there is an endogenous margin
between employment and nonemployment and labor market imperfections (such as search,
bargaining, or effi
ciency wages) that create a wedge between wages and outside options. In
both cases the automation decisions of fi rms fail to internalize the gap between the market wage and the opportunity cost of labor, leading to excessive automation (see Acemoglu and Restrepo 2018a).
226 Daron Acemoglu and Pascual Restrepo
which implies that the total
number of workers earning rents declines with
automation.
Moreover, the appendix shows that (gross) output is now given by
I N +1
J I
N J
K
L
L
L
(12)
Y = B
A
A
,
I
N + 1
J
I
N
J
and GDP is still given by Y – RK. Equation (12) highlights that there is now a misallocation of labor across tasks—output can be increased by allocating
more workers to tasks ( I, J ) where their marginal product is greater (because of the rents they are earning).
Equation (12) further implies that the impact of automation on GDP is
given by
d GDP
W (1 + )
R(1
)
dK
dL
= ln
ln
R
+
W
A
.
dI
( I )
( I )
dI
dI
L
M
Excessive
Excessive displacement
Productivity effect>0
automation<0
of labor<0
The new term W( dL / dI ) captures the fi rst- order losses from a decline in A
employment in tasks ( I, J ). These losses arise because by automating jobs where workers earn rents, fi rms are eff ectively displacing workers to other
tasks in which they have a lower marginal product and earn a strictly lower
wage, which increases the extent of misallocation.
The point highlighted here is much more general. Without labor market
frictions, automation increases GDP (and net output), so at the very least
it is possible to redistribute the gains that it creates to make workers—of
diff erent skill levels—better off . Labor market frictions change this picture.
In the presence of such frictions, fi rms’ automation decisions do not inter-
nalize the fact that the marginal product of labor is above its opportunity
cost, or equivalently, do not recognize that there are fi rst- order losses that
workers will suff er as a result of automation. Consequently, equilibrium
automation could reduce GDP and welfare and there may not be a way
to make (all) workers better off , even with tools for costless redistribution.
Under these circumstances, a utilitarian planner would choose a lower level
of automation than the equilibrium.15
8.6 Concluding
Remarks
Despite the growing concerns and intensifying debate about the implica-
tions of automation for the future of work, the economics profession and
popular discussions lack a satisfactory conceptual framework. To us this
15. Naturally, if the planner could remove the rents, or the labor market frictions underpinning them, then the equilibrium would be restored to effi
ciency. Nevertheless, most sources of
rents, including search, bargaining, and effi
ciency wages, would be present in the constrained
effi
cient allocations as well.
Artifi cial Intelligence, Automation, and Work 227
lack of appropriate conceptual approach is also the key reason why much
of the debate is characterized by a false dichotomy between the view that
automation will spell the end of work for humans and the argument that
technologies will always tend to increase the demand for labor as they have
done in the past.
In this chapter, we summarized a conceptual framework that can help
understand the implications of automation and bridge the opposite sides
of this false dichotomy. At the center of our framework is a task- based
approach, where automation is conceptualized as replacing labor in tasks
that it used to perform. This type of replacement causes a direct displace-
ment eff ect, reducing labor demand. If this displacement eff ect is not coun-
terbalanced by other economic forces, it will reduce labor demand, wages,
and employment. But our framework also emphasizes that there are several
countervailing forces. These include the fact that automation will reduce the
costs of production and thus create a productivity eff ect, the induced capital
accumulation, and the deepening of automation—technological advances
that increase the productivity of machines in tasks that have already been
automated.
Our framework also emphasizes that these countervailing forces are gen-
erally insuffi
cient to totally balance out the implications of automation. In
particular, even if these forces are strong, the displacement eff ect of automa-
tion tends to cause a decline in the share of labor in national income. But
we know from the history of technology and industrial development that
despite several waves of rapid automation, the growth process has been more
or less balanced, with no secular downward trend in the share of labor in
national income. We argue this is because of another powerful force: the
creation of new tasks in which labor has a comparative advantage, which
fosters a countervailing reinstatement eff ect for labor. These tasks increase
the demand for labor and tend to raise the labor share. When they go hand-
in-hand with automation, the growth process is balanced and it need not
imply a dismal scenario for labor.
Nevertheless, the adjustment process is likely to be slower and more pain-
ful than this account of balance between automation and new tasks at fi rst
suggests. This is because the reallocation of labor from its existing jobs
and tasks to new ones is a slow process, in part owing to time- consuming
search and other labor market imperfections. But even more ominously, new
tasks require new skills. When the education sector does not keep up with
the demand for new skills, the mismatch between skills and technologies is
bound to complicate the adjustment process and hinder the productivity
gains from new technologies.
Our framework further suggests that there are additional reasons for the
productivity slowdown. At the center of these is a tendency for excessive
automation because of the tax treatment of capital investments and labor
market imperfections. Excessive automation directly reduces productivity,
228 Daron Acemoglu and Pascual Restrepo
but may have even more powerful indirect eff ects because it redirects tech-
nological improvements away from productivity- enhancing activities that
lead to the creation of new tasks to excessive eff orts at the extensive margin
of automation, a picture that receives informal support from the current
singular focus on AI and deep learning.
We would like to conclude by pointing out a number of additional issues
that may be important in understanding the full impact of AI and other auto-
mation technologies on future prospects of labor. We believe that these issues
can be studied using simple extensions of the framework presented here.
First, we have emphasized the role of the productivity eff ect in partially
counterbalancing the displacement eff ect created by automation. However,
this counterv
ailing eff ect works by increasing the demand for products. As
we have also seen, automation tends to increase inequality. If, as a conse-
quence of this distributional impact, the rise in real incomes resulting from
automation ends up in the hands of a narrow segment of the population
with much lower marginal propensity to consume than those losing incomes
and their jobs, these countervailing forces would be weakened and might
operate much more slowly. This imbalance in the distribution of the gains
from automation might slow down the creation of new tasks as well.
Second, our analysis highlighted the negative consequences of a short-
age of skills for realizing the productivity gains from automation and for
inequality. In practice, the problem may be workers acquiring the wrong
types of skills rather than a general lack of skills. For example, if AI and
other new automation technologies necessitate a mix of numeracy, com-
munication, and problem- solving skills diff erent than those emphasized in
current curricula, this would have implications similar to those of a shortage
of skills, but it cannot be overcome by just increasing educational spending
with current educational practices remaining intact. One important con-
sideration in this respect is that there is little concrete information about
what types of skills new technologies will complement, underscoring the
importance of further empirical work in this area.
Third, government policies and labor market institutions may impact not
just the speed of automation (and thus whether there is excessive auto-
mation), but what types of technologies will receive more investments. To
the extent that some uses of AI may complement labor more or generate
opportunities for more rapid creation of new tasks, an understanding of
the impact of various policies, including support for academic and applied
research, and social factors on the path of development of AI is critical.
Last but not least, the development and adoption of technologies that re-
instate labor cannot be taken for granted. If we do not fi nd a way of creating
shared prosperity from the productivity gains generated by new technolo-
gies, there is a danger that the political reaction to these technologies may
slow down or even completely stop their adoption and development. This
Artifi cial Intelligence, Automation, and Work 229
underscores the importance of studying the distributional implications of
AI and robotics, the political economy reactions to it, and the design of
new and improved institutions for creating more broadly shared gains from