The Economics of Artificial Intelligence
Page 37
advantage in higher- indexed tasks.7
The threshold I denotes the frontier of automation possibilities: it
describes the range of tasks that can be automated using current available
technologies in AI, industrial robots, various computer- assisted technolo-
gies, and other forms of “smart machines.”
We also simplify the discussion by assuming that both the supply of labor,
L, and the supply of machines, K, are fi xed and inelastic. The fact that the supply of labor is inelastic implies that changes in labor demand impact the
share of labor in national income and the wage, but not the level of employ-
ment. We outline below how this framework can be easily generalized to
accommodate changes in employment and unemployment.
8.3.2 Types of Technological Change
Our framework incorporates four diff erent types of technological ad-
vances. All advances increase productivity, but as we will see with a very
diff erent impact on the demand for labor and wages.
1. Labor- augmenting technological advances: Standard approaches in
macroeconomics and labor economics typically focus on labor- augmenting
technological advances. Such technological changes correspond to increases
(or perhaps an equi- proportionate increase) in the function ( x) . Our anal-
L
ysis will show that they are in fact quite special, and the implications of auto-
mation and AI are generally very diff erent from those of labor- augmenting
advances.
2. Automation (at the extensive margin): We consider automation to be an
expansion of the set of tasks that are technologically automated as repre-
sented by the parameter I.
7. Our theoretical framework builds on Zeira (1998) who develops a model where fi rms
produce intermediates using labor- intensive or capital- intensive technologies. Zeira focuses on how wages aff ect the adoption of capital- intensive production methods and how this margin amplifi es productivity diff erences across countries and over time. In contrast, we focus on the implications of automation—modeled here as an increase in the set of tasks that can be produced by machines, represented by I—for the demand for labor, wages, and employment, and we also study the implications of the introduction of new tasks. In Acemoglu and Restrepo (2016), we generalize Zeira’s framework in a number of other dimensions and also endogenize the development of automation technologies and new tasks.
Artifi cial Intelligence, Automation, and Work 213
3. Deepening of automation (or automation at the intensive margin): An-
other dimension of advances in AI and robotics technology will tend to
increase the productivity of machines in tasks that are already automated,
for example, by replacing existing machines with newer, more productive
vintages. In terms of our model, this corresponds to an increase in the ( x)
M
function for tasks x < I. We will see that this type of deepening of automation has very diff erent implications for labor demand than automation (at
the extensive margin).
4. Creation of new tasks: As emphasized in Acemoglu and Restrepo
(2016), another important aspect of technological change is the creation of
new tasks and activities in which labor has a comparative advantage. In our
model this can be captured in the simplest possible way by an increase in N.
8.3.3 Equilibrium
Throughout, we denote the equilibrium wage rate by W and the equilib-
rium cost of machines (or the rental rate) by R. An equilibrium requires
fi rms to choose the cost- minimizing way of producing each task and labor
and capital markets to clear.
To simplify the discussion, we impose the following assumption
( N )
( I )
(A1)
L
> W > L
.
( N
1)
R
( I )
M
M
The second inequality implies that all tasks in [ N – 1, I ] will be produced by machines. The fi rst inequality implies that the introduction of new
tasks—an increase in N—will increase aggregate output. This assumption
is imposed on the wage- to-rental rate ratio, which is an endogenous object;
the appendix provides a condition on the stock of capital and labor that is
equivalent to this assumption (see assumption [A2]).
We also show in the appendix that aggregate output (GDP) in the equi-
librium takes the form
I N +1
N I
K
L
(3)
Y = B
,
I
N + 1
N
I
where
I
N
(4)
B = exp
ln
( x) dx + ln
( x) dx .
M
L
N 1
I
Aggregate output is given by a Cobb- Douglas aggregate of the capital stock
and employment. This resulting aggregate production function in equation
(3) is itself derived from the allocation of the two factors of production to
tasks. More important, the exponents of capital and labor in this production
function depend on the extent of automation, I, and the creation of new
tasks, as captured by N.
Central to our focus is not only the impact of new technologies on pro-
214 Daron Acemoglu and Pascual Restrepo
ductivity, but also on the demand for labor. The appendix shows that the
demand for labor can be expressed as
Y
(5)
W = ( N
I )
.
L
This equation is intuitive in view of the Cobb- Douglas production func-
tion in equation(3), since it shows that the wage (the marginal product of
labor) is equal to the average product of labor—which we will also refer to
as “productivity”—times the exponent of labor in the aggregate production
function.
Equation (5) implies that the share of labor in national income is given by
(6)
s = WL = N
I.
L
Y
8.4 Technology and Labor Demand
8.4.1 The Displacement Eff ect
Our fi rst result shows that automation (at the extensive margin) indeed
creates a displacement eff ect, reducing labor demand as emphasized in sec-
tion 8.2, but also that it is counteracted by a productivity eff ect, pushing
toward greater labor demand.
Specifi cally, from equation (5) we directly obtain
d ln W
d ln( N
I )
d ln( Y / L)
(7)
=
+
.
dI
dI
dI
Displacement effect <0
Productivity effect > 0
Without the productivity eff ect, automation would always reduce labor
demand because it is directly replacing labor in tasks that were previously
performed by workers. Indeed, if the productivity eff ect is limited, automa-
tion will reduce labor demand and wages.
8.4.2 Counteracting the Displacement Eff ect I: The Productivity Eff ects
The productivity eff ect, on the other hand, captures the important idea
that by increasing productivity, automation raises labor demand in the tasks
that are not automated. As highlighted in the previous section, there are two
complementary manifestations of the productivity eff ect. The fi rst works
by increasing the demand for labor in nonautomated tasks in the industries
where automation is ongoing. The second works by raising the demand
for labor in other industries. The productivity eff ect shown in equation (7)
combines these two mechanisms.
One important implication of the decomposition in equation (7) is that, in
Artifi cial Intelligence, Automation, and Work 215
contrast to some popular discussions, the new AI and robotics technologies
that are more likely to reduce the demand for labor are not those that are
brilliant and highly productive, but those that are “so- so”—just productive
enough to be adopted but not much more productive or cost- saving than
the production processes that they are replacing. Interestingly, and related
to our discussion on missing productivity, if new automation technologies
are so-so, they would not bring major improvements in productivity either.
To elaborate further on this point and to understand the productivity
implications of automation technologies better, let us also express the pro-
ductivity eff ect in terms of the physical productivities of labor and machines
and factor prices as follows:
d ln( Y / L)
W
R
= ln
> 0.
dI
( I )
ln
( I )
L
M
The fact that this expression is positive, and that new automation technolo-
gies will be adopted, follows from assumption (A1). Using this expression,
the overall impact on labor demand can be alternatively written as
d ln W
W
R
(8)
=
1
+ ln
.
dI
N
I
( I )
ln
( I )
L
M
Displacement effect <0
Productivity effect > 0
This expression clarifi es that the displacement eff ect of automation will
dominate the productivity eff ect and thus reduce labor demand (and wages)
when ( I )/ R ≈ ( I ) / W, which is exactly the case when new technologies M
L
are so-so—only marginally better than labor at newly automated tasks. In
contrast, when ( I )/ R >> ( I ) / W , automation will increase productivity M
L
suffi
ciently to raise the demand for labor and wages.
Turning next to the implications of automation for the labor share, equa-
tion (6) implies
ds
(9)
L = 1 < 0,
dI
so that regardless of the magnitude of the productivity eff ect, automation
always reduces the share of labor in national income. This negative impact
on the labor share is a direct consequence of the fact that automation always
increases productivity more than the wage, d ln( Y / L) / d I > d ln W / dI (itself directly following from equation [7], which shows that the impact on wages
is given by the impact on productivity minus the displacement eff ect).
The implications of standard labor- augmenting technological change,
which corresponds to a (marginal) shift-up of the ( x) schedule, are very
L
diff erent from those of automation. Labor-augmenting technologies leave
the form of the wage equation (5) unchanged, and increase average output
216 Daron Acemoglu and Pascual Restrepo
per worker, Y /L, and the equilibrium wage, W, proportionately, and thus do not impact the share of labor in national income.8
8.4.3 Counteracting the Displacement Eff ect II: Capital Accumulation
We have so far emphasized the displacement eff ect created by new auto-
mation technologies. We have also seen that the productivity eff ect counter-
acts the displacement eff ects to some degree. In this and the next subsection,
we discuss two additional countervailing forces.
The fi rst force is capital accumulation. The analysis so far assumed that
the economy has a fi xed supply of capital that could be devoted to new
machines (automation technologies). As a result, a further increase in auto-
mation (at the extensive margin) increases the demand for capital and thus
the equilibrium rental rate, R. This may be understood as the short- run
eff ect of automation.
Instead, we may envisage the “medium- run” eff ect as the impact of
these technologies after the supply of machines used in newly automated
tasks expands as well. Because machines and labor are q- complements, an
increase in the capital stock, with the level of employment held constant at
L, increases the real wage and reduces the rental rate. Equation (8) shows
that this change in factor prices makes the productivity eff ect more powerful
and the impact on the wage more likely to be positive.
In the limit, if capital accumulation fi xes the rental rate at a constant
level (which will be the case, for example, when we have a representative
household with exponential discounting and time- separable preferences),
the productivity eff ect will always dominate the displacement eff ect.9
Crucially, however, equation (6) still applies, and thus automation contin-
ues to reduce the labor share, even after the adjustment of the capital stock.
8.4.4 Counteracting the Displacement Eff ect III:
Deepening of Automation
Another potentially powerful force counteracting the displacement eff ect
from automation at the extensive margin comes from the deepening of auto-
mation (or automation at the intensive margin), for example, because of
improvements in the performance of already- existing automation technolo-
8. A small shift-up of ( x) does not violate assumption (A1) because at the margin it was L
strictly cost- saving to use machines. A larger labor- augmenting technological change may result in a violation of assumption (A1). At this point, only tasks below an endogenous threshold I< I would be automated, and labor- augmenting technologies could also reduce I, increasing the labor share in national income.
9. Assuming that production exhibits constant returns to scale, the productivity gains from any technology accrue to both capital and labor. In particular, for any constant returns to scale production function, we have d ln Y |
= s d ln W + (1 – s ) d ln R, where d ln Y |
> 0 denotes
K,L
L
L
K,L
the productivity gains brought by technology holding the use of capital and labor constant, and s is the labor share. If the rental rate is constant in the long run, then d ln R = 0 and all L
productivity gains accrue to the relatively inelastic factor, labor.
Artifi cial Intelligence, Automation, and Work 217
gies or the replacement of such technologies with newer, more productive
vintages. This increase in the productivity of machines in tasks that are
already automated corresponds in our model to an increase in the function
( x) in tasks below I.
M
To explore the implications of this type of change in the simplest possible
way, let us suppose that ( x) = i
n all automated tasks, and consider an
M
M
increase in the productivity of machines by d ln > 0, with no change in
M
the extensive margin of automation, I. The implications of this change in
the productivity of machines on equilibrium wages and productivity can
be obtained as
d ln W = d ln Y / L = I
N + 1
(
) d ln > 0.
M
Hence, deepening of automation will tend to increase labor demand and
wages, further counteracting the displacement eff ect. Note, however, that
as with capital accumulation, in our model this has no impact on the share
of labor in national income, as can be seen from the fact that wages and
productivity increase by exactly the same amount.
8.4.5 New Tasks and the Comparative Advantage of Labor
Much more powerful than the countervailing eff ects of capital accumula-
tion and the deepening of automation is the creation of new tasks in which
labor has a comparative advantage. These tasks include both new, more
complex versions of existing tasks and the creation of new activities, which
are made possible by advances in technology. In terms of our framework,
they correspond to increases in N.
An increase in N—the creation of new tasks—raises productivity by
d ln Y / L
R
W
= ln
ln
> 0,
dN
( N
1)
( N )
M
L
which is positive from assumption (A1).
More important for our focus here, the creation of new tasks also in-
creases labor demand and equilibrium wages by creating a reinstatement
eff ect counter acting the displacement eff ect. In particular,
d ln W
R
W
(10)
= ln
ln
+
1
.
dN
( n 1)
( N )
N
I
M
L
Reinstatement effect>0
Productivity effect>0
In contrast to capital accumulation and the deepening of automation,
which increase the demand for labor but do not aff ect the labor share, equa-
tion (6) implies that new tasks increase the labor share, that is,
ds
L = 1.
dN
218 Daron Acemoglu and Pascual Restrepo
The centrality of new tasks can be understood when viewed from a com-
plementary historical angle. Automation is not a recent phenomenon. As we