The Economics of Artificial Intelligence
Page 36
equilibrating force could be powerful enough to make the growth process
balanced.
2. Some automation technology platforms, especially AI, may facilitate
the creation of new tasks. A recent report by Accenture identifi ed entirely
new categories of jobs that are emerging in fi rms using AI as part of their
production process (Accenture PLC 2017). These jobs include “trainers” (to
train the AI systems), “explainers” (to communicate and explain the output
of AI systems to customers), and “sustainers” (to monitor the performance
of AI systems, including their adherence to prevailing ethical standards).
The applications of AI to education, health care, and design may also
result in employment opportunities for new workers. Take education. Exist-
Artifi cial Intelligence, Automation, and Work 207
ing evidence suggests that many students, not least those with certain learn-
ing disabilities, will benefi t from individualized education programs and
personalized instruction (Kolb 1984). With current technology, it is pro-
hibitively costly to provide such services to more than a small fraction of
students. Applications of AI may enable the educational system to become
more customized, and in the process create more jobs for education profes-
sionals to monitor, design, and implement individualized education pro-
grams. Similar prospects exist in health care and elderly care services.
8.2.4 Revisiting the False Dichotomy
The conceptual framework outlined above, which will be further elabo-
rated in the next section, clarifi es why the current debate is centered on a false
dichotomy between disastrous and totally benign eff ects of automation.
Our task- based framework underscores that automation will always
create a displacement eff ect. Unless neutralized by the countervailing forces,
this displacement eff ect could reduce labor demand, wages, and employ-
ment. At the very least, this displacement eff ect implies that a falling share
of output will accrue to labor. These possibilities push against the benign
accounts emphasizing that technology always increases the demand for
labor and benefi ts workers.
Our framework does not support the alarmist perspectives stressing the
disastrous eff ects of automation for labor either. Rather, it highlights several
countervailing forces that soften the impact of automation on labor. More
important, as we have argued in the previous subsection, the creation of new
labor- intensive tasks has been a critical part of the adjustment process in the
face of rapid automation. The creation of new tasks is not just manna from
heaven. There are good reasons why market incentives will endogenously
lead to the creation of new tasks that gain strength when automation itself
becomes more intensive. Also, some of the most defi ning automation tech-
nologies of our age, such as AI, may create a platform for the creation of
new sets of tasks and jobs.
At the root of some of the alarmism is the belief that AI will have very dif-
ferent consequences for labor than previous waves of technological change.
Our framework highlights that the past is also replete with automation
technologies displacing workers, but this need not have disastrous eff ects
for labor. Nor is it technologically likely that AI will replace labor in all or
almost all of the tasks in which it currently specializes. This limited remit of
AI can be best understood by contrasting the current nature and ambitions
of AI with those of its fi rst coming under the auspices of “cybernetics.” The
intellectual luminaries of cybernetics, such as Norbert Wiener, envisaged
the production of Human- Level Artifi cial Intelligence—computer systems
capable of thinking in a way that could not be distinguished from human
intelligence—replicating all human thought processes and faculties (Nilsson
2009). In 1965, Herbert Simon predicted that “machines will be capable,
208 Daron Acemoglu and Pascual Restrepo
within twenty years, of doing any work a man can do” (Simon 1965, 96).
Marvin Minsky agreed, declaring in 1967 that “Within a generation, I am
convinced, few compartments of intellect will remain outside the machine’s
realm” (Minsky 1967, 2).
Current practice in the fi eld of AI, especially in its most popular and prom-
ising forms based on deep learning and various other “big data” methods
applied to unstructured data, eschews these initial ambitions and aims at
developing applied artifi cial intelligence—commercial systems specializing
in clearly delineated tasks related to prediction, decision- making, logistics,
and pattern recognition (Nilsson 2009). Though many occupations involve
such tasks—and so AI is likely to have a displacement eff ect in these tasks—
there are still many human skills that we still cannot automate, including
complex reasoning, judgment, analogy- based learning, abstract problem-
solving, and a mixture of physical activity, empathy, and communication
skills. This reading of the current practice of AI suggests that the potential
for AI and related technological advances to automate a vast set of tasks
is limited.
8.2.5 Flies in the Ointment
Our framework so far has emphasized two key ideas. First, automation
does create a potential negative impact on labor through the displacement
eff ect and also by reducing the share of labor in national income. But sec-
ond, it can be counterbalanced by the creation of new tasks (as well as the
productivity eff ect, capital accumulation and the deepening of automation,
which tend to increase the demand for labor, even though they do not gener-
ally restore the share of labor in national income to its preautomation levels).
The picture we have painted underplays some of the challenges of adjust-
ment, however. The economic adjustment following rapid automation can
be more painful than the process we have outlined for a number of reasons.
Most straightforward, automation changes the nature of existing jobs,
and the reallocation of workers from existing jobs and tasks to new ones
is a complex and often slow process. It takes time for workers to fi nd new
jobs and tasks in which they can be productive, and periods during which
workers are laid off from their existing jobs can create a depressed local or
national labor market, further increasing the costs of adjustment. These
eff ects are visible in recent studies that have focused on the adjustment of
local US labor markets to negative demand shocks, such as Autor, Dorn,
and Hanson (2013), who study the slow and highly incomplete adjustment
of local labor markets in response to the surge in Chinese exports, Mian
and Sufi (2014), who investigate the implications of the collapse in housing
prices on consumption and local employment, and perhaps more closely
related to our focus, Acemoglu and Restrepo (2017), who fi nd employment
and wage declines in areas most exposed to one specifi c type of automation,
the introduction of industrial robots in manufacturing.
Artifi cial Intelligence, Automation, and Work 209
The historical record also un
derscores the painful nature of the adjust-
ment. The rapid introduction of new technologies during the British Indus-
trial Revolution ultimately led to rising labor demand and wages, but this
was only after a protracted period of stagnant wages, expanding poverty,
and harsh living conditions. During an eighty- year period extending from
the beginning of the Industrial Revolution to the middle of the nineteenth
century, wages stagnated and the labor share fell, even as technological
advances and productivity growth were ongoing in the British economy,
a phenomenon which Allen (2009) dubs the “Engel’s pause” (previously
referred to as the “living standards paradox”; see Mokyr [1990]).
There should thus be no presumption that the adjustment to the changed
labor market brought about by rapid automation will be a seamless, costless,
and rapid process.
8.2.6 Mismatch between Skills and Technologies
It is perhaps telling that wages started growing in the nineteenth- century
British economy only after mass schooling and other investments in human
capital expanded the skills of the workforce. Similarly, the adjustment to
the large supply of labor freed from agriculture in early twentieth- century
America may have been greatly aided by the “high school movement,” which
increased the human capital of the new generation of American workers
(Goldin and Katz 2010). The forces at work here are likely to be more general
than these examples. New tasks tend to require new skills. But to the extent
that the workforce does not possess those skills, the adjustment process will
be hampered. Even more ominously, if the educational system is not up to
providing those skills (and if we are not even aware of the types of new skills
that will be required so as to enable investments in them), the adjustment
will be greatly impeded. Even the most optimistic observers ought to be
concerned about the ability of the current US educational system to identify
and provide such skills.
At stake here is not only the speed of adjustment, but potential produc-
tivity gains from new technologies. If certain skills are complementary to
new technologies, their absence will imply that the productivity of these
new technologies will be lower than otherwise. Thus the mismatch between
skills and technologies not only slows down the adjustment of employment
and wages, but holds back potential productivity gains. This is particularly
true for the creation of new tasks. The fact that while there is heightened
concerns about job losses from automation, many employers are unable to
fi nd workers with the right skills for their jobs underscores the importance
of these considerations (Deloitte and the Manufacturing Institute 2011).
8.2.7 Missing Productivity and Excessive Automation
The issues raised in the previous subsection are important not least because
a deep puzzle in any discussion of the impact of new technologies is miss-
210 Daron Acemoglu and Pascual Restrepo
ing productivity growth—the fact that while so many sophisticated tech-
nologies are being adopted, productivity growth has been slow. As pointed
out by Gordon (2016), US productivity growth since 1974 (with the excep-
tion of the period from 1995 to 2004) compares dismally to its postwar per-
formance. While the annual rate of labor productivity growth of the US
economy averaged 2.7 percent between 1947 and 1973, it only averaged
1.5 percent between 1974 and 1994. Average productivity growth rebounded
to 2.8 percent between 1995 and 2004, and then fell again to only 1.3 percent
between 2005 and 2015 (Syverson 2017). How can we make sense of this?
One line of attack argues that there is plenty of productivity growth, but it
is being mismeasured. But, as pointed out by Syverson (2017), the pervasive
nature of this slow down, and the fact that it is even more severe in industries
that have made greater investments in information technology (Acemoglu
et al. 2014), make the productivity mismeasurement hypothesis unlikely to
account for all of the slowdown.
Our conceptual framework suggests some possible explanations. They
center around the possibility of “excessive automation,” meaning faster
automation than socially desirable (Acemoglu and Restrepo 2016, 2018a).
Excessive automation not only creates direct ineffi
ciencies, but may also hold
productivity growth down by wastefully using resources and displacing labor.
There are two broad reasons for excessive automation, both of which we
believe to be important. The fi rst is related to the biases in the US tax code,
which subsidizes capital relative to labor. This subsidy takes the form of
several diff erent provisions, including additional taxes and costs employ-
ers have to pay for labor, subsidies in the form of tax credits and acceler-
ated depreciation for capital outlays, and additional tax credit for interest
rate deductions in case of debt- fi nanced investments (AEI 2008; Tuzel
and Zhang 2017). All of these distortions imply that at the margin, when
a utilitarian social planner would be indiff erent between capital and labor,
the market would have an incentive to use machines, giving an ineffi
cient
boost to automation. This ineffi
ciency could translate into slow productivity
growth because the substitution of labor for machines worsens the misal-
location of capital and labor.
Even absent such a fi scal bias, there are natural reasons for excessive auto-
mation. Labor market imperfections and frictions also tend to imply that
the equilibrium wage is above the social opportunity cost of labor. Thus
a social planner would use a lower shadow wage in deciding whether to
automate a task than the market, creating another force toward excessive
automation. The implications of this type of excessive automation would
again include slower productivity growth than otherwise.
Finally, it is possible that automation has continued at its historical pace,
or may have even accelerated recently, but the dismal productivity growth
Artifi cial Intelligence, Automation, and Work 211
performance we are witnessing is driven by a slowdown in the creation of
new tasks or investment in other productivity- enhancing technologies (see
Acemoglu and Restrepo 2016). A deceleration in the creation of new tasks
and technologies other than automation would also explain why the period
of slow productivity growth coincided with poor labor market outcomes,
including stagnant median wages and a decline in the labor share.
There are natural reasons why too much emphasis on automation may
come at the cost of investments in other technologies, including the creation
of new tasks. For instance, in a setting where technologies are developed
endogenously using a common set of resources (e.g., scientists), there is a
natural trade- off between faster automation and investments in other types
of technologies (Acemoglu and Restrepo 2016). Though it is at the moment
impossible to know whether the redirection of research resources away from
the creation of new tasks and toward autom
ation has played an important
role in the productivity slowdown, the almost singular focus in the corporate
sector and research community on AI, applications of deep learning, and
other big data methods to automate various tasks makes it at least plausible
that there may be too much attention devoted to automation at the expense
of other technological breakthroughs.
8.3 A Model of Automation, Tasks, and the Demand for Labor
In the previous section, we provided an intuitive discussion of how auto-
mation in general, and robotics and AI in particular, is expected to impact
productivity and the demand for labor. In this section, we outline a for-
mal framework that underlines these conclusions. Our presentation will be
somewhat informal and without any derivations, which are all collected in
the appendix.
8.3.1 A Task- Based Framework
We start with a simplifi ed version of the task- based framework intro-
duced in Acemoglu and Restrepo (2016). Aggregate output is produced by
combining the services of a unit measure of tasks x ∈ [ N – 1, N] according to the following Cobb- Douglas (unit elastic) aggregator
N
(1)
ln Y =
ln y( x) dx,
N 1
where Y denotes aggregate output and y( x) is the output of task x. The fact that tasks run between N – 1 and N enables us to consider changes in the range of tasks, for example, because of the introduction of new tasks,
without altering the total measure of tasks in the economy.
Each task can be produced by human labor, ℓ( x), or by machines, m( x), depending on whether it has been (technologically) automated or not. In
212 Daron Acemoglu and Pascual Restrepo
particular, tasks x ∈ [ N – 1, I] are technologically automated, so can be produced by either labor or machines, while the rest are not technologically
automated, so must be produced with labor:
( x) ( x) +
( x) m( x) if x
N
1, I
M
(2)
y( x) =
L
( x) ( x)
if x
( I, N .
L
Here, ( x) is the productivity of labor in task x and is assumed to be increas-L
ing, while ( x) is the productivity of machines in automated tasks. We
M
assume that ( x)/ ( x) is increasing in x, and thus labor has a comparative L
M