by Ajay Agrawal
medium run, aff ecting workers across many professions and skill levels.1 The
magnitude of these disruptions will depend on two important factors: the
speed and the factor bias of progress in AI.
On the fi rst factor, measured productivity has increased rather slowly
in recent years, even as the world seems to be captured by AI fever.2 If AI-
related innovations enter the economy at the same slow pace as suggested
by recent productivity statistics, then the transition will be slower than, for
example, the wave of mechanization in the 1950– 1970s, and the resulting dis-
ruptions may not be very signifi cant. However, there are three possible alter-
natives: First, some suggest that productivity is signifi cantly undermeasured,
for example, because quality improvements are not accurately captured. The
best available estimates suggest that this problem is limited to a few tenths
of a percentage point (see, e.g., the discussion in Groshen et al. [2017]).
Furthermore, there are also unmeasured deteriorations in productivity, for
example, declines in service quality as customer service is increasingly auto-
mated. Second, the aggregate implications of progress in AI may follow a
delayed pattern, similar to what happened after the introduction of comput-
ers in the 1980s. Robert Solow (1987) famously quipped that “you can see the
computer age everywhere but in the productivity statistics.” It was not until
the 1990s that a signifi cant rise in aggregate productivity could be detected,
after sustained investment in computers and a reorganization of business
practices had taken place. Third, it is of course possible that a signifi cant
discontinuity in productivity growth occurs, as suggested, for example, by
proponents of a technological singularity (see, e.g., Kurzweil 2005).
On the second factor, the disruptions generated by AI- related innova-
tions depend on whether they are labor- saving, using the terminology of
Hicks (1932), that is, whether at a given wage the innovations lead to less
demand for labor. Some suggest that artifi cial intelligence will mainly assist
humans in being more productive, and refer to such new technologies as
intelligence- assisting innovation (IA), rather than AI. Although we agree
that most AI- related innovations are likely to be complementary to at least
some jobs, we believe that in taking a broader perspective, progress in AI
1. For example, Frey and Osborne (2017) warn that 47 percent of jobs in the US economy are at risk of being automated by advances in AI- related fi elds. Areas in which human intelligence has recently become inferior to artifi cial intelligence include many applications of radiology, trading in fi nancial markets, paralegal work, underwriting, driving, and so forth.
2. For example, Google Trends reveals that search interest in the topic “artifi cial intelligence”
has quadrupled over the past four years.
AI and Its Implications for Income Distribution and Unemployment 351
is more likely to substitute for human labor, or even to replace workers
outright, as we will assume in some of our formal models below.
We believe that the primary economic challenge posed by the proliferation
of AI will be one of income distribution. We economists set ourselves too
easy a goal if we just say that technological progress can make everybody
better off —we also have to say how we can make this happen. This chapter is an attempt to do so by discussing some of the key economic research issues
that this raises.3
In section 14.2 of this chapter, we provide a general taxonomy of the
relationship between technological progress and welfare. We fi rst observe
that in a truly fi rst- best economy—in which complete risk markets are avail-
able before a veil of ignorance about innovations is lifted—all individuals
will share in the benefi ts of technological progress. However, since the real
world does not correspond to this ideal, redistribution is generally needed to
ensure that technological progress generates Pareto improvements. If mar-
kets are perfect and redistribution is costless, it can always be ensured that
technological progress makes everybody better off . The same result holds
if the costs of redistribution are suffi
ciently low. In all these cases, there
can be political unanimity about the desirability of technological progress.
However, if redistribution is too costly, it may be impossible to compen-
sate the losers of technological progress, and they will rationally oppose
progress. Even worse, if the economy suff ers from market imperfections,
technological progress may actually move the Pareto frontier inwards, that
is, some individuals may necessarily be worse off . Finally, we observe that
the fi rst welfare theorem does not apply to the process of innovation, and as
a result, privately optimal innovation choices may move the Pareto frontier
inward.
In section 14.3, we decompose the mechanisms through which innovation
leads to inequality into two channels. First, inequality rises because innova-
tors earn a surplus. Unless markets for innovation are fully contestable, the
surplus earned by innovators is generally in excess of the costs of innovation
and includes what we call innovator rents. We discuss policies that aff ect the
sharing of such rents, such as antitrust policies and changes in intellectual
property rights. The second channel is that innovations aff ect market prices;
they change the demand for factors such as diff erent types of labor and
capital, which aff ects their prices and generates redistributions. For example,
AI may reduce a wide range of human wages and generate a redistribution
to entrepreneurs. From the perspective of our fi rst- best benchmark with
complete insurance markets, these factor price changes represent pecuni-
ary externalities. We discuss policies to counter the eff ects of the resulting
factor price changes.
3. An important, and maybe even more diffi
cult, complementary question, which is beyond
the scope of this chapter, is to analyze the political issues involved.
352 Anton Korinek and Joseph E. Stiglitz
In section 14.4, we develop a simple formal model of worker- replacing
technological change, that is, we introduce a machine technology that acts
as a perfect substitute for human labor. We study the implications for wages
and discuss policy remedies. In the short run, an additional unit of machine
labor that is added to the economy earns its marginal product, but also gen-
erates a zero- sum redistribution from labor to traditional capital because
it changes the relative supply of the two. In the long run, the machine tech-
nology turns labor into a reproducible factor. Thus, in the long run, growth
will likely be limited by some other irreproducible factor, and all the benefi ts
of technological progress will accrue to that factor. However, since it is in
fi xed supply, it can be taxed and the proceeds can be redistributed without
creating distortions. Hence a Pareto improvement is easily achieved.
In a second model, we demonstrate how changes in patent length and
capital taxation can act as a second- best device to redistribute if lump sum
&n
bsp; transfers between workers and innovators are not available. A longer patent
life both delays how quickly innovations enter the public domain, lower-
ing consumer prices, and increases the incentives of innovators to produce
worker- replacing machines. However, the resulting losses for workers can
be made up for by imposing a distortionary tax on capital and providing
transfers, so long as the supply elasticity of capital is suffi
ciently low.
We also discuss the implications of endogenous factor bias in technologi-
cal change. Worker- replacing technological progress should make capital-
saving innovations more desirable, providing some relief to workers. We
also note that our economy is developing more and more into a service
economy, and that the large role of government in many service sectors
(e.g., education, healthcare, etc.) creates ample scope for interventions to
support workers.
In section 14.5, we observe two categories of reasons for why innova-
tion may lead to technological unemployment. The fi rst category of reasons
arises because wages cannot adjust, even in the long run: effi
ciency wage
theory implies that employers may fi nd it effi
cient to pay wages above the
market- clearing level so that workers have incentives to exert proper eff ort.
If technological progress lowers the marginal product of workers, and hence
their real wage declines below their cost of living, then classic nutritional
effi
ciency wage theories apply: unemployment would result because (in the
absence of government support) workers could not survive working for the
market- clearing wage and it would pay employers to raise real wages above
the market- clearing level because of the resulting increase in worker pro-
ductivity. The second category of technological unemployment arises as a
transition phenomenon, when jobs are replaced at a faster rate than workers
can fi nd new ones. We discuss a variety of factors that may slow down the
adjustment process. Effi
ciency wage arguments may also play an important
role as a transitional phenomenon, in particular if workers’ notion of fair
wages is sticky. Finally, we discuss that jobs may not only provide wages but
AI and Its Implications for Income Distribution and Unemployment 353
also meaning and note that, unless societal attitudes change with the pro-
liferation of AI, it may be welfare enhancing to subsidize jobs rather than
simply redistributing resources.
In section 14.6, we take a longer- term perspective that is somewhat more
speculative and discuss the potential implications of superhuman artifi cial
intelligence. We consider two scenarios: one in which some humans use tech-
nology to enhance themselves and attain superhuman intelligence, and one
in which autonomous machines that are completely separate from humans
reach superhuman intelligence. In both cases, the superior productivity of
superior intelligence will likely lead to vast increases in income inequality.
From a Malthusian perspective, the superintelligent entities are likely to
command a growing share of the scarce resources in the economy, creating
the risk of pushing regular humans below their subsistence level. We discuss
corrective actions that could be taken.
14.2 Technological Progress and Welfare: A Taxonomy
In 1930 Keynes wrote an essay on the “Economic Possibilities of our
Grandchildren,” in which he described how technological possibilities may
translate into utility possibilities. He worried about the quality of life that
would emerge in a world with excess leisure. And he thought all individuals
might face that quandary. But what has happened in recent years has raised
another possibility: innovation could lead to a few very rich individuals—
who may face this challenge—whereas the vast majority of ordinary workers
may be left behind, with wages far below what they were at the peak of the
industrial age.
So let us start by considering the arrival of a new technology that par-
tially (or fully) replaces workers and let us ask the question: would their
standard of living necessarily decline? We will consider this question in a
number of diff erent settings, providing a taxonomy for how technological
progress might aff ect the welfare of diff erent groups in society depending
on the environment.
14.2.1 First
Best
We start with a fi rst- best scenario in which we assume that all markets
are perfect: this includes risk markets that are free of adverse incentive
eff ects and that allow individuals to insure against the advent of innovations
“behind the veil of ignorance,” that is, before they know whether they will
be workers or innovators. The main purpose for considering this idealized
setting is to demonstrate that from an ex ante perspective, compensating
workers for the losses imposed by technological progress is a question of
economic effi
ciency not redistribution.
If risk markets were perfect and accessible to all agents before they knew
their place in the economy, then all agents would be insured against any risk
354 Anton Korinek and Joseph E. Stiglitz
that might aff ect their well- being, including the risk of innovation reduc-
ing the value of their factor endowment. For example, workers would be
insured against the risk of declining wages.4 This leads us to the following
observation:
Observation 1) Consider a fi rst- best world in which all individuals have
access to a perfect insurance market “behind the veil of ignorance,” that is,
before they know whether they will be innovators or workers. If an innovation
occurs in such a world, the winners would compensate the losers as a matter of
optimal risk sharing. As a result, technological progress always makes every-
body better off , and there is political unanimity in supporting it.
This is a powerful observation because it reminds us that if we had an ideal
market, something that very much looks like redistribution would naturally
emerge. In our fi rst- best economy, there are no losers from technological
progress. Losers only exist if risk markets are imperfect compared to this
benchmark. In more technical language, worker- replacing technological
progress imposes pecuniary externalities on workers, which lead to inef-
fi ciency when risk markets are imperfect (see, e.g., Stiglitz and Weiss 1981;
Greenwald and Stiglitz 1986; Geanakoplos and Polemarchakis 1986; or
more recently Dávila and Korinek 2018).
This implies that policy measures to mitigate or undo the pecuniary exter-
nalities arising from technological progress—for example, redistribution
programs—make the economy’s allocation more effi
cient from an ex ante
perspective, rather than “interfering” with economic effi
ciency. They bring
us closer to the allocation that a well- functioning risk market would achieve.
Policymakers who oppose redistribution to compensate the losers of innova-
tion because it interferes with the free market seem to—inappropriately, in
and after individuals know their place in the economy. Even though they
may pretend to preach about idealized free markets, they clearly have not
understood the full implications of how an idealized free market would
work, that is, that such a market would provide precisely the type of insur-
ance that they are opposing.
In practice, workers who might be replaced by technological progress can-
not purchase insurance contracts against being replaced, so in the absence
of adequate government assistance, they are in fact hurt by the innovation.
Of course there are good reasons for why such idealized risk markets are
not present in the real world.
First, the limited lifespan of humans makes it diffi
cult to write insur-
ance contracts that stretch over multiple generations. Workers would have
had to obtain the described insurance a long time ago, before AI was well
4. We will discuss the reasons why this is typically not the case in practice below.
AI and Its Implications for Income Distribution and Unemployment 355
conceived and its implications were clear, when the associated insurance pre-
mium would have been commensurately low. Perhaps their farsighted ances-
tors could have written state- contingent contracts on their behalf. Today,
obtaining insurance against AI- reducing wages would require workers to
pay large amounts since the possibility is very real. In short, eff ective insur-
ance would have had to take place behind a “veil of ignorance” about the
likely advent of AI.
To put it another way, in this perspective the fi rst “insurable damage”
to the individual occurs at the time that the probability of an innovation
becomes nonnegligible, for at that time the insurance premium required
for income smoothing becomes signifi cant, and her welfare is lowered. The
individual would have wanted to buy insurance against the risk that her
insurance premium would go up. Thus, in a perfect market, insurance mar-
kets would have to go back at least to a date at which there was a negligible
probability that the innovation occurs. This presents a problem: it may be
that at the moment that the concept of AI is formulated precisely enough
to be an insurable event (and therefore becomes an insurable event), it has