by Ajay Agrawal
a non- zero probability.
Second, even for more limited time periods, risk markets with respect to
technological change are clearly not perfect. Among the main reasons are
information problems.
Describing the State Space. This starts with the basic problem alluded to
earlier of how diffi
cult it is to describe the future state space.5 We cannot
easily write a contract on something before it has been invented. Address-
ing this problem would require that an individual has to be insured against
any technological event that leads to lower wages. But any such insurance
contract would necessarily have adverse incentive eff ects.
More broadly, the curse of asymmetric information that inhibits insur-
ance markets is as prevalent here as it is elsewhere.
Adverse Selection. Innovation leads to important adverse- selection prob-
lems. Some people in the market are more informed than others. In an ideal
market, the winners of innovation would provide insurance to the losers, and
the winners (e.g., entrepreneurs) would almost certainly be better informed
than the losers (e.g., workers).
Moral Hazard. Innovation may also be subject to moral hazard problems,
that is, the presence of insurance may aff ect the likelihood that the insured
event occurs. Although workers are unlikely to aff ect the pace of innova-
tion in AI, the actions of innovators may be, to some extent, aff ected. If
they were to completely insure away all their returns from innovation, there
5. Interestingly, this type of information problem is easy to deal with after innovation has occurred, because then we know what has been invented and in which state we are, but very diffi
cult to capture in ex ante contracts.
356 Anton Korinek and Joseph E. Stiglitz
would be scant incentive to exert eff ort.6 Since, in a perfect insurance world,
the winners would insure the losers, full insurance would lead to stagnation.
Insurance and Redistribution
A natural counterpart to observation 1 is that in the absence of perfect
insurance markets “behind the veil of ignorance,” ensuring that progress
leads to a Pareto improvement generally requires redistribution. If workers
have access to some insurance against the risk of AI but not perfect insur-
ance, this does not remove the need for redistributions.
For example, obtaining AI insurance today would require workers to pay
a large premium. Of course, conceptually, if one went back in time, before
AI was well conceived and its implications clear, one might argue that the
premium would be low. But even that might not be so, since premiums for
large events, even with small probability, can be high. In any case, at the very
moment of conception of AI—the fi rst possible moment that one could
conceivably have written a policy—AI would still have distributional con-
sequences; workers would have to pay a premium to divest themselves of
this risk, and thus they would be worse off compared to the innovators, the
winners.
14.2.2 Perfect Markets Ex Post and No Costs of Redistribution
Our next case pertains to a world that may be described as a second- best
world without the perfect insurance markets referred to earlier, but in which,
ex post, all markets are functioning well and there can be costless redistribu-
tions. This case covers several critical results that, although obvious at some
level, often get lost in the debate about AI and technological progress more
generally.
Observation 2) If redistribution is costless and appropriate redistribu-
tions are made, then technological progress is always desirable for all agents.
In that case, there is political unanimity in supporting technological progress.
For convenience, and in conformity to conventional usage, we will refer
the world with costless redistribution but otherwise perfectly functioning
markets, as fi rst best ex post; though we remind the reader that the previous
analysis suggested that in a true fi rst- best, workers would have insurance
against the risks from AI, such that they would commensurately share any
gains from innovation. If the world is fi rst- best ex post in the sense thus
6. Some might argue that this problem is equally hard to deal with before or after innovation has occurred. If we tax innovators ex post, it destroys incentives just as much as if we fully insure away all returns from innovation. However, in both cases, the signifi cance of any adverse eff ects is not clear. Innovators are at least partially driven by non- pecuniary motives. And partial insurance or partial redistribution are always an option. If Bill Gates had been told, ex ante, that government would take away 50 percent of his returns over $10 billion, there is little reason to believe that it would have had any signifi cant eff ect on innovation and investment.
Ex post, taxing the winners in “winner takes all” games may have only small incentive eff ects.
AI and Its Implications for Income Distribution and Unemployment 357
Fig. 14.1 Pareto frontier before and after innovation with costless redistribution
defi ned, then the utility possibilities curve (or Pareto frontier) moves out.
We provide an example in fi gure 14.1, which depicts a utility possibilities
frontier for two types of agents, workers, and entrepreneurs. In the example,
technological progress increases the maximum utility level of entrepreneurs
for any given level of utility of workers.7 Innovation has increased produc-
tion possibilities, and with lump sum redistributions, an expansion in pro-
duction possibilities automatically implies an expansion in utility possibili-
ties, that is, that everybody could be better off .
The fact that they could be better off does not mean that they will be better off . That depends on institutional arrangements. In fi gure 14.1, we denote
the initial equilibrium by E and the after- innovation equilibrium by E . We 0
1
have deliberately not called it a competitive market equilibrium: markets do
not exist in a vacuum (see e.g. Stiglitz et al. 2015). They are structured by
rules and regulations, for example, concerning intellectual property rights
and antitrust policies, and there may be tax and other policies in place. We
thus simply refer to E and E as the before and after innovation (institution-0
1
given) equilibrium given the existing set of institutions. Note that in the
example drawn in the fi gure, workers are worse off . That would normally
be the case with what Hicks referred to as labor- saving innovations, that
is, innovations that at a given wage lead to less demand for labor. Artifi cial
intelligence appears to be a labor- saving innovation. In the simple formal
models of worker- replacing innovations that we work out below, that is
clearly the case.
This in turn has two important implications.
First, without adequate redistribution it makes sense for workers to resist
the innovation. Luddism—the movement named after the possibly fi ctional
7. More generally, we could defi ne a multidimensional utility possibilities frontier by adding any number of categories of individuals, or even naming the individuals.
358 Anton Korinek and Joseph E. Stiglitz
character Ned Ludd that opposed automation in the textile sector in the late
eighteenth and early nineteenth centuries in England—is a rational response
for workers who are worse off from automation and who are not suffi
ciently
compensated.
Second, in a democracy in which workers are in a majority it would make
sense for enlightened innovators to support redistribution, to make sure
that workers are at least not worse off . With redistribution, both innovators
and workers can be better off . If appropriate redistribution is made so that
everybody shares in the fruits of technological progress, there will again be
political unanimity in supporting technological progress—progress will not
be politically contentious.
There might be signifi cant debate about how much compensation workers
should receive, that is where in the “northeast corner of E ” society should
0
be. On the one hand, this debate concerns the distribution of the surplus gen-
erated by innovation. On the other hand, labor- saving innovation reduces
wages, which generates a redistribution from workers to other factor owners
like rentiers and capitalists, for which workers may seek compensation. This
redistribution represents a pecuniary externality from the innovation, as we
will discuss in further detail in section 14.3.2.
In fi gure 14.1, we have marked in bold that part of the postinnovation
Pareto frontier that represents a Pareto improvement and lies to the north-
east of E . A range of philosophical principles can be adduced for determin-
0
ing what is a “just” division of the fruits of innovation. Behavioral econom-
ics may provide insights into what kinds of divisions might be acceptable.8
Of course, the innovation may not be labor saving, and the equilibrium
E itself could be to the northeast of E . Although this case is easier, the dis-1
0
tribution of the gains from innovation and any associated pecuniary exter-
nalities and rents may still be contentious, especially if they lead to large
disparities in income. Distributive issues can also interact with production,
as emphasized for example, by the effi
ciency wage theories that we consider
in greater depth in section 14.5.
14.2.3 Perfect Markets but Costly Redistribution
There is another possibility—that as we try to redistribute, the new utility
possibility curve may lie inside of the old utility possibilities frontier near the
original equilibrium. This may be the case even in a world that is fi rst- best
8. Consider a model in which workers and innovators have to agree on whether the innovation is acceptable. The innovator has the power to set the division of the gains (i.e., where along the curve Northeast of E the new equilibrium lies), but the workers have the power to accept or 0
reject. This is the standard ultimatum game, for which there is a large body of literature suggesting that at least some of the fruits of innovation have to be shared with workers. If they perceive the allocation of benefi ts to be unfair, they would rather be worse off (e.g., at the original point without the innovation) than at the point that just makes them indiff erent to where they were before. See Fehr and Schmidt (2003).
AI and Its Implications for Income Distribution and Unemployment 359
ex post, that is, in which all the conditions for Pareto effi
ciency would be
satisfi ed ex post after the innovation has taken place.
Observation 3) If the world is fi rst- best ex-post, but redistribution is lim-
ited or costly, then a Pareto improvement may not be possible, and some groups
in society may oppose technological progress. With a suffi
ciently inequality-
averse social welfare function, societal welfare may be reduced.
This case is illustrated in fi gure 14.2. The utility possibilities frontier is
constrained by the costs imposed by redistribution. Even though it might
appear that innovation could make everyone better off technologically, given
the existing set of institutions of that economy, it actually can’t—there may
not be scope for avoiding utility losses for workers.
Some economists argue that the world looks like fi gure 14.2, and that if
we try to transfer from innovators to workers, so much output is lost that
workers are still worse off . If that is the case, then one cannot say that the
innovation is a Pareto improvement. One hesitates to use the word “innova-
tion.” It is a change, perhaps a technological change, which has had the eff ect
of making some people better off and others worse off . It is a distribution-
inducing change and will be contentious.
A social welfare function that places no weight on inequality—which
treats a dollar to rich innovators the same as a dollar to a poor worker—
would, of course, conclude that the innovation is desirable. But with a more
natural, inequality- averse social welfare function, the so-called innovation
is welfare decreasing.
The workers who lose out would rationally oppose the innovation. If
workers are in a majority and innovators wish to maintain their position, it
would behoove the innovators to think harder about how to engage in redis-
tribution. This is, of course, a collective action problem for innovators—for
individual innovators, the contribution to economy- wide inequality is typi-
Fig. 14.2 Potential Pareto frontier with costly redistribution
360 Anton Korinek and Joseph E. Stiglitz
cally limited, even if their collective behavior makes workers worse off . As
a result, innovators often devote eff ort to actions that enhance their market
power—lowering real incomes of workers still further—and to not paying
taxes (both via clever tax avoidance using the existing legal framework, and
via political lobbying to provide special exemptions from taxation for their
industries). Disregarding, in our view unwisely, that their actions may stir
up political opposition to innovation, some innovators go further and argue
for weakening the progressivity of the tax system and a smaller state, so there
are less public resources to provide for the well- being of the workers who
are hurt by innovation.
According to a long- run version of “trickle- down” economics, repeated
innovations will eventually increase the wealth of innovators so much that
the benefi ts will trickle down to regular workers. In this view, a Pareto
improvement is always possible in the long run, as in fi gure 14.1, even if
an entire generation of workers is hurt in the short- to-medium run. This
is a possibility and, in fact, the fi rst industrial revolution may be an ex-
ample. During the Industrial Revolution, workers eventually obtained
enough human capital—which was publicly provided, as is in the interests
of the innovators—so that the wages of almost all increased. In the cur-
rent context, however, once machines are smart enough, innovators may
no longer have incentives to support the public fi nancing of human capital
accumulation, and it may well be that workers’ standards of living decrease.
In particular, in a political system dominated by money, the innovators,
increasingly rich, may use their economic and political infl uence to resist
redistribution. Furthermore, even if l
ong- run trickle- down economics was
correct, it may lead to tremendous suff ering and social upheaval in the short
run. It may also—understandably—not be very credible if innovators prom-
ise that once they are rich enough, they will support workers, but that they
are not quite rich enough yet.
This leads to the important question: How costly is redistribution in prac-
tice? As we noted earlier, markets do not exist in a vacuum. They are struc-
tured by laws and regulations and by how those laws and regulations are
enforced. The outcome is the so-called “market” distribution of income,
which is then subject to taxes and transfer, leading to an after- tax distribu-
tion of income. But this conventional distinction may not be quite accurate:
the rules of the game concerning redistribution aff ect the market income
distribution, and are themselves endogenous, aff ected by the rules of the
political game, which in turn are aff ected by the distribution of wealth. (See
Piketty, Saez, and Stantcheva (2014) and Stiglitz 2017.) The points that we
have denoted E and E describe the initial outcome and the outcome after-0
1
technological change, assuming that laws, regulations, institutions, and so forth
remain unchanged. But, of course, it is not reasonable to expect that they
would remain unchanged with the advent of a change as signifi cant as AI.
Setting aside the endogeneity of the rules themselves, each set of (feasible)
AI and Its Implications for Income Distribution and Unemployment 361
laws, regulations, institutions, and so on defi nes a feasible utility possibilities
frontier. We can think of the second- best utility possibilities frontier as the
outer envelope of all these frontiers. As the outer envelope, the second- best
utility possibilities frontier provides more fl exibility for redistribution than
does that associated with one particular set of rules, regulations, and institu-
tions. This refl ects that any changes in laws, regulations, or institutions, and
so forth will also have redistributive eff ects. Given this additional fl exibility,
the likelihood that a Pareto improvement as in fi gure 14.1 can be achieved
is greater. We provide further arguments for why this is likely to be the case