by Ajay Agrawal
wages have to fall or that other complementary factors like capital have to
adjust enough for labor market equilibrium to be restored at or above the
historic wage.
14.5.1 Effi
ciency Wage Theory and Nonadjustment of Wages
The fi rst category of technological unemployment arises when wages do
not adjust for structural reasons. Effi
ciency wage theory emphasizes that
productivity depends on wages and so employers may have reasons to pay
wages above the market- clearing level. The original effi
ciency wage paper
(Stiglitz 1969) noted one of the reasons for this: that income disparities can
weaken worker morale. Akerlof and Yellen (1990) have formalized this into
the “fair wage hypothesis.”
If fairness considerations are signifi cant enough, and workers think that a
decrease in their wages is “unfair” (e.g., because the income of entrepreneurs
increases so entrepreneurs could easily “aff ord” pay increases), it means that
the scope of labor- saving progress that shifts the utility possibilities curve
out without redistributions is very limited. Similar results hold if workers’
well- being and eff orts are related to relative incomes. The new utility pos-
sibilities curve may lie outside the old one to the “north” of E , that is, there 0
is scope for a Pareto improvement in principle; but it may lie inside of the old
utility possibilities curve near E , that is, the utility possibilities of workers 1
decrease for a given level of utility of entrepreneurs because workers reduce
their eff ort so much that the eff ective labor supply declines—any gains from
technology are more than off set by increased shirking. Shapiro and Stiglitz
(1984) emphasize that paying a wage above the market- clearing level reduces
shirking, leading to unemployment.
An even more daunting example of effi
ciency wages may arise if automa-
tion continues and the marginal product of labor for low- skill workers falls
below their cost of living at what they view as their basic subsistence (even
if they exert their best eff ort). Unless basic social services are provided to
such workers, a nutritional effi
ciency wage model applies in that case, similar
to what Stiglitz (1976) described for developing countries: employers could
not pay a market- clearing wage because they know that this would be insuf-
fi cient for their employees to provide for themselves and remain productive.29
We will follow up on this theme in the fi nal section of our chapter.
In traditional effi
ciency wage models, the unemployment eff ects of effi
-
29. Even worse outcomes could emerge in the presence of imperfect capital markets, if
expenditures on health and nutrition at one date aff ect productivity at later dates.
AI and Its Implications for Income Distribution and Unemployment 379
ciency wages are permanent, part of the long- run equilibrium. For example,
if technological change leads to greater inequality (or better information
about the existing level of inequality), morale eff ects and the resulting
effi
ciency wage responses imply that the equilibrium level of unemploy-
ment rises.
However, effi
ciency wage arguments may also contribute to slowing down
the transition to a new equilibrium after an innovation, as we will explore
subsection 14.5.2.
Minimum Wages and Nonadjustment of Wages
An alternative reason why wages may not adjust to the market- clearing
level are minimum wage laws. Basic economics implies that there will be
unemployment if wages are set to an excessive level. Although this is a
theoretical possibility, recent experience in the United States has repeatedly
shown that modest increases in minimum wages from current levels have
hardly any employment eff ects but raise the income of minimum wage work-
ers, which may have positive aggregate demand eff ects since low- income
workers have a high marginal propensity to consume (see, e.g., Schmitt
2013). From an economic theory perspective, these observations are possible
because wages are not determined in a purely Walrasian manner—there is
a signifi cant amount of bargaining involved when prospective employers
and employees match—and increases in minimum wages substitute for the
lacking bargaining power of workers (see, e.g., Manning 2011).
14.5.2 Technological Unemployment as a Transition Phenomenon
The second category of technological unemployment is as a transition
phenomenon, that is, when technological change makes workers redundant
at a faster pace than they can fi nd new jobs or that new jobs are created. This
phenomenon was already observed by Keynes (1932). It is well understood
that there is always a certain “natural” or “equilibrium” level of unemploy-
ment as a result of churning in the labor market. In benchmark models of
search and matching to characterize this equilibrium level of unemployment
(see Mortensen and Pissarides 1994, 1998), employment relationships are
separated at random, and workers and employers need to search for new
matches to replace them. The random shocks in this framework can be
viewed as capturing, in reduced form, phenomena such as life cycle transi-
tions but also technological progress in individual fi rms. In this view, an
increase in the pace of technological progress corresponds to a higher job
separation rate and results in a higher equilibrium level of unemployment.
The transition may be especially prolonged if technology implies that the
old skills of workers become obsolete and they need to acquire new skills
and/or fi nd out what new jobs match their skills (see, e.g., Restrepo 2015).
Even if in the long run workers adjusted to AI, the transition may be
diffi
cult. Artifi cial intelligence will impact some sectors more than others,
380 Anton Korinek and Joseph E. Stiglitz
and there will be signifi cant job dislocation. As a general lesson, markets on
their own are not good at structural transformation. Often, the pace of job
destruction is greater than the pace of job creation, especially as a result of
imperfections in capital markets, inhibiting the ability of entrepreneurs to
exploit quickly new opportunities as they are opened up.
The Great Depression as an Example of Transitional Unemployment
The Great Depression can be viewed as being caused by rapid pace of
innovation in agriculture (see Delli Gatti et al. 2012a). Fewer workers were
needed to produce the food that individuals demanded, resulting in marked
decline in agriculture prices and income, leading to a decline in demand for
urban products. In the late 1920s, these eff ects became so large that long-
standing migration patterns were reversed.
What might have been a Pareto improvement turned out to be an immis-
erizing technological change, as both those in the urban and rural sector
suff ered.
The general result is that noted earlier: with mobility frictions and rigidi-
ties (themselves partly caused be capital market imperfections, as workers in
t
he rural sector couldn’t obtain funds to obtain the human capital required
in the urban sector and to relocate) technological change can be welfare
decreasing. The economy can be caught, for an extensive period of time,
in a low- level equilibrium trap, with high unemployment and low output.
In the case of the Great Depression, government intervention (as a
by-product of World War II) eventually enabled a successful structural
transformation: the intervention was not only a Keynesian stimulus, but
facilitated the move from rural farming areas to the cities where manu-
facturing was occurring at the time and facilitated the retraining of the
labor force, helping workers acquire the skills necessary for success in an
urban manufacturing environment, which were quite diff erent from those
that ensured success in a rural, farming environment. It was, in this sense,
an example of a successful industrial policy.
There are clear parallels to the situation today in that a signifi cant fraction
of the workforce may not have the skills required to succeed in the age of AI.
Transitional Effi
ciency Wage Theory
Effi
ciency wage arguments may also slow down the transition to a new
equilibrium after technological progress. For example, if worker morale
depends on last period’s wages, it may be diffi
cult to reduce wages to the
market- clearing level after a labor- saving innovation, and unemployment
may persist for a long time.30
30. In the limiting case, employers may simply keep wages fi xed to avoid negative morale eff ects, and unemployment would persist forever—or until some off setting shock occurs.
AI and Its Implications for Income Distribution and Unemployment 381
14.5.3 Jobs and Meaning
The potentially widespread destruction of jobs can have large human
consequences that go beyond just economics because jobs provide not only
income but also other mental services such as meaning, dignity, and fulfi ll-
ment to humans. Whether this is a legacy of our past, and whether individu-
als could fi nd meaning in other forms of activities, mental or physical, is a
matter of philosophical debate.
If workers derive a separate benefi t from work in the form of meaning,
then job subsidies are a better way of ensuring that technological advances
are welfare enhancing than simply providing lump sum grants (e.g., through
the provision of a universal basic income), as some are suggesting in
response to the inequalities created by AI.
This discussion is, of course, a departure from the usual neoclassical for-
mulation, where work only enters negatively into individual’s well- being.
There are some that claim that individuals’ deriving dignity and meaning
from work is an artifact of a world with labor scarcity. In a workerless AI
world, individuals will have to get their identity and dignity elsewhere, for
example, through spiritual or cultural values. The fact that most humans can
fi nd a meaningful life after retirement perhaps suggests that there are good
substitutes for jobs in providing meaning.
14.6 Longer- Term Perspectives: AI and the Return of Malthus?
There is a fi nal point that is worth discussing in a chapter on the implica-
tions of artifi cial intelligence for inequality. This point relates to a somewhat
longer- term perspective. Currently, artifi cial intelligence is at the stage where
it strictly dominates human intelligence in a number of specifi c areas, for
instance playing chess or Go, identifying patterns in x-rays, driving, and so
forth. This is commonly termed narrow artifi cial intelligence. By contrast,
humans are able to apply their intelligence across a wide range of domains.
This capacity is termed general intelligence.
If AI reaches and surpasses human levels of general intelligence, a set of
radically diff erent considerations apply. Some techno- optimists predict the
advent of general artifi cial intelligence for as early as 2029 (see Kurzweil
2005), although the median estimate in the AI expert community is around
2040 to 2050, with most AI experts assigning a 90 percent probability to
human- level general artifi cial intelligence arising within the current century
(see Bostrom 2014). A minority believes that general artifi cial intelligence
will never arrive. However, if human- level artifi cial general intelligence is
reached, there is broad agreement that AI would soon after become super-
intelligent, that is, more intelligent than humans, since technological pro-
gress would likely accelerate, aided by the intelligent machines. Given these
382 Anton Korinek and Joseph E. Stiglitz
predictions, we have to think seriously about the implications of artifi cial
general intelligence for humanity and, in the context of this chapter, for what
it implies for our economy as well as for inequality.
Assuming that our social and economic system will be maintained upon
the advent of artifi cial general intelligence and superintelligence,31 there
are two main scenarios. One scenario is that man and machine will merge,
that is, that humans will “enhance” themselves with ever more advanced
technology so that their physical and mental capabilities are increasingly
determined by the state of the art in technology and AI rather than by
traditional human biology (see, e.g., Kurzweil, 2005). The second scenario
is that artifi cially intelligent entities will develop separately from humans,
with their own objectives and behavior (see, e.g., Bostrom 2014; Tegmark
2017). As we will argue below, it is plausible that the two scenarios might
diff er only in the short run.
First Scenario: Human Enhancement and Inequality
The scenario that humans will enhance themselves with machines may
lead to massive increases in human inequality, unless policymakers recog-
nize the threat and take steps to equalize access to human enhancement tech-
nologies.32 Human intelligence is currently distributed within a fairly narrow
range compared to the distance between the intelligence of humans and that
of the next- closest species. If intelligence becomes a matter of ability to pay,
it is conceivable that the wealthiest (enhanced) humans will become orders
of magnitude more productive—“more intelligent”—than the unenhanced,
leaving the majority of the population further and further behind. In fact,
if intelligence enhancement becomes possible, then—unless preemptive
actions are taken—it is diffi
cult to imagine how to avoid such a dynamic.
For those who can aff ord it, the incentive to purchase enhancements is great,
especially since they are in competition with other wealthy humans who may
otherwise leapfrog them. This is even more so in an economy which is, or
is perceived to be, a winner- take- all economy and/or in which well- being
is based on relative income. Those who cannot aff ord the latest technology
will have to rely on what is in the public domain, and if the pace of innova-
tion increases, the gap between the best technology and what is publicly
available will increase.
A useful analogy is to compare human enhancement tech
nology to health
care—technology to maintain rather than enhance the human body. Dif-
31. Researchers who work on the topic of AI safety point out that there is also a risk of doomsday scenarios in which a suffi
ciently advanced artifi cial intelligence eradicates humanity
because humans stand in the way of its goals. See, for example, Bostrom (2014) who elaborates on this using the example of a “paperclip maximizer”—an AI that has been programmed to produce as many paperclips as possible, without regard for other human goals, and who realizes that humans contain valuable raw materials that should better be transformed into paperclips.
32. In many respects, the issues are parallel to those associated with performance- enhancing drugs. In sports, these have been strictly regulated, but in other arenas, they have not.
AI and Its Implications for Income Distribution and Unemployment 383
ferent countries have chosen signifi cantly diff erent models for how to provide
access to health care, with some regarding it as a basic human right and
others allocating it more according to ability to pay. In the United States, for
example, the expected life spans of the poor and the wealthy have diverged
signifi cantly in recent decades, in part because of unequal access to health
care and ever more costly new technologies that are only available to those
who can pay. The diff erences are even starker if we look at humanity across
nations, with the expected life span in the richest countries being two- thirds
longer than in the least developed countries (see, e.g., UN 2015). Like with
health care, it is conceivable that diff erent societies will make signifi cantly
diff erent choices about access to human enhancement technologies.
Once the wealthiest enhanced humans have separated suffi
ciently far from
the unenhanced, they can eff ectively be considered as a separate species of
artifi cially intelligent agents. To emphasize the diff erence in productivities,
Yuval Harari (2017) has dubbed the two classes that may result “the gods”
and “the useless.” In that case, the long- run implications of our fi rst scenario
coincide with the second scenario.
Second Scenario: Artifi cially Intelligent Agents and the Return of Malthus
We thus turn to the scenario that artifi cially intelligent entities develop
separately from regular (or unenhanced) humans. One of the likely char-