The Economics of Artificial Intelligence

by Ajay Agrawal

  so that she cannot recoup the investment into an innovation.15

  13. This approach relies on individuals or companies that are willing to innovate in exchange for nonpecuniary rewards such as prestige or, alternatively, on a calculated decision that providing free technology will steer potential customers or employees toward an innovator’s platform, as seems to be the case in the fi eld of AI.

  14. For discussions of the merits of alternative ways of funding and incentivizing innovation, see Dosi and Stiglitz (2014), Baker, Jayadev, and Stiglitz (2017), Stiglitz (2008), and Korinek and Ng (2017).

  15. See Stiglitz (1987b) and Dasgupta and Stiglitz (1988). When the number of fi rms is limited and there is Cournot competition, there will also be rents associated with innovation. For more general theoretical discussions of industrial structure and innovation, see Dasgupta and Stiglitz (1980a 1980b) and Stiglitz and Greenwald (2015, ch. 5).

  AI and Its Implications for Income Distribution and Unemployment 367

  Second, if innovative activity is contestable, that is, if there is a suffi

  -

  ciently large set of potential innovators with equal skills, then the expected

  rents to innovative activity are competed down to zero, that is, the marginal

  entrant into innovative activity is indiff erent between innovating or not.16

  However, given that the payoff s to innovation are highly stochastic, there

  will be winners and losers ex post. In the context of new technologies, the

  distribution of payoff s seems to be increasingly skewed, with a small num-

  ber of entrepreneurs earning gigantic payoff s and the vast majority earning

  little in return for their eff orts. This gives rise to signifi cant inequality even

  among innovators.17

  In either case, the returns earned by an innovator may not correspond

  closely to the social returns to the innovation; in particular, some of the

  returns may refl ect the capture of profi ts that would otherwise have gone to

  other entrepreneurs.

  Policies to Share the Surplus of Innovators

  There is a growing consensus that one of the sources of the growth of

  inequality is the growth of rents, including the rents that innovators earn

  in excess of the cost of innovation (see e.g. Korinek and Ng 2018). Taxing

  and redistributing such rents has an important role in ensuring that AI and

  other advances in technology are Pareto improving. Also, anti- trust policies

  may lower such rents, ensuring that the benefi ts of innovations are more

  widely shared, as more competition lowers consumer prices from which all

  benefi t. From the perspective of low wage workers who lose from innova-

  tion, targeted expenditure programs fi nanced by high rent taxes may be

  of greater benefi t than the lowering of prices, the benefi t of which will go

  disproportionately to those who have high spending power.

  Moreover, changes in intellectual property rights (IPRs) aff ect who

  receives the benefi ts of innovation—and thus the “incidence” of innova-

  tion, since IPRs are instrumental in providing extended market power to

  innovators.18

  Additionally, public research—with government or the public at large

  appropriating the returns, rather than allowing private fi rms to do so—

  together with stronger competition policies, might reduce the scope for

  16. Given the diffi

  culty of predicting the success of innovative activity or of even assigning

  success probabilities, it is questionable how effi

  ciently this mechanism works in practice. For

  example, there may be excessive entry because of overoptimism by some potential entrepreneurs, or there may be insuffi

  cient entry because of imperfect insurance markets for risk averse

  entrepreneurs. If some are better at innovating than others (and know it), then these individuals will enjoy inframarginal innovation rents (on average.)

  17. If there are diminishing returns to the allocation of resources to innovation, the effi cient

  resource allocation will entail there being rents associated with innovation. Who captures these rents will be aff ected by the institutional (including tax) structure. In the absence of adequate mechanisms for the public capturing those rents, there may be excessive investments in innovation, a standard result in the economics of the commons.

  18. Especially when there is Bertrand competition, the benefi ts of innovation may be quickly shared with consumers upon the termination of patents.

  368 Anton Korinek and Joseph E. Stiglitz

  monopolies capturing large fractions of the returns to innovation, and thus

  enhance the likelihood that AI will be Pareto improving.

  Workers may also note that the innovations that ultimately led to AI—

  including those created by private entrepreneurs—build on signifi cant

  public support. Society as a whole, but not necessarily this generation of

  innovators, paid for this knowledge, and should therefore share in the sur-

  plus generated by the innovation. One proposal to ensure that workers share

  in the benefi ts of innovation—and are less likely to lose from it—is to give

  workers shares in enterprises to ensure that their welfare goes up in tandem

  with that of shareholders/ innovators as a whole.

  14.3.2 Eff ects on Others

  Innovation also leads to large redistributions among others in the

  economy who are not directly involved in the process of innovation, for ex-

  ample, workers who experience a sudden increase or decline in the demand

  for their labor. These redistributions can thus be viewed as externalities from

  innovation, and they are one of the main reasons why innovation raises

  concerns about inequality. We distinguish two categories of such externali-

  ties, pecuniary and nonpecuniary externalities. We discuss both in detail in

  the following.

  Pecuniary Externalities: Price and Wage Changes

  Among the most prominent implications of technological change is that it

  aff ects the prices of factors of production (including wages) and of produced

  goods. Hicks (1932) already observed that innovations generally change the

  demand for factors and will, in equilibrium, lead to factor price changes,

  especially changes in wages. The price and wage changes that result from

  innovations represent pecuniary externalities. Traditional general equilib-

  rium theory, following Arrow and Debreu, emphasized that pecuniary exter-

  nalities are fully consistent with Pareto effi

  ciency. However, the benchmark

  of Pareto effi

  ciency is blind to the distribution of income. Even if the equi-

  librium reached after an innovation is Pareto effi

  cient, the pecuniary exter-

  nalities lead to redistributions and imply that there are winners and losers.19

  If—as many technologists predict—artifi cial intelligence directly replaces

  human labor, the demand for human labor will go down, and so will wages.

  More generally, innovations typically reduce demand for specifi c types of

  labor with specifi c human capital. For example, self- driving cars will likely

  depress the wages of drivers, or radiology- reading AI may lower the wages

  of traditional radiologists. Conversely, AI has certainly led to an increase in

  demand for computer scientists and has greatly increased their wages, in par-

  19. Greenwald and Stiglitz (1986) and Geanakoplos and Polemarchakis (1986) demonstrated tha
t pecuniary externalities also matter for effi

  ciency when there are market imperfections such

  as imperfect information and incomplete markets; market equilibrium will as a result not even be Pareto effi

  cient.

  AI and Its Implications for Income Distribution and Unemployment 369

  ticular in subfi elds that are directly related to AI. Since AI is a general pur-

  pose technology, there are reasons to believe that advances in AI will rever-

  berate throughout many diff erent sectors and lead to signifi cant changes in

  wages throughout the economy in coming decades. Similar arguments can

  be made about the demand for and the value of diff erent types of specifi c

  capital, as well as the demand for and prices of particular products.

  Even though there are frequently losers, technological progress by defi -

  nition shifts out the production possibilities frontier. This implies that the

  total dollar gain of the winners of progress exceeds the dollar loss of the

  losers.20 In section 14.4 below, we will use this property of technological pro-

  gress to argue that under relatively broad conditions this should enable the

  redistribution that is necessary to ensure that innovation leads to a Pareto

  improvement: the gains that arise to some factor owners as a result of tech-

  nological progress are excess returns that are like unearned rents and could

  be taxed away without introducing distortions into the economy.

  Although we noted that pecuniary externalities are generally viewed as

  Pareto effi

  cient, there are two reasons for why they are likely to be associated

  with ineffi

  ciency in practice. First, Greenwald and Stiglitz (1986) and Geana-

  koplos and Polemarchakis (1986) demonstrated that pecuniary externalities

  matter for effi

  ciency when there are market imperfections such as imperfect

  information and incomplete markets; market equilibrium will as a result

  not be Pareto effi

  cient. Compared to the benchmark of idealized insurance

  markets “behind the veil of ignorance” that we discussed in section 14.2, the

  pecuniary externalities from innovation are clearly ineffi

  cient. Additional

  market imperfections are likely to lead to additional ineffi

  ciencies. Second,

  if the pecuniary externalities from innovation give rise to the need for redis-

  tributive policies that are costly to perform, the policy response will generate

  additional ineffi

  ciencies.

  Policies to Counter Wage Declines

  Aside from lump sum transfers, there are a range of further policies to

  counter the wage declines that are experienced by workers who are dis-

  placed by machines, even for low- skill jobs. These include wage subsidies

  and earned income tax credits. If bargaining power in labor markets is biased

  toward employers, an increased minimum wage can also help ensure that no

  one who works full time is in poverty. Furthermore, ensuring high aggregate

  demand—and thus a low unemployment rate—also increases the bargain-

  ing power of workers and leads to higher wages.

  Other policies aimed at increasing the demand for especially low- skill

  labor include any measures that raise the wages of workers that are substi-

  tutes, for example, higher wages in the public sector as well as an increase in

  20. If lump sum transfers were feasible, the winners could compensate the losers. However, in the absence of such compensation, social welfare may be lower.

  370 Anton Korinek and Joseph E. Stiglitz

  public investments and other public expenditures; all of these policies help

  to drive up wages in the economy more generally.

  Policies that could be used to fi nance such measures include carbon taxes,

  which would encourage resource- saving innovation at the expense of labor-

  saving innovation. It would thus simultaneously address two of most serious

  global problems, global climate change and inequality.21

  Furthermore, the elimination of tax deduction for interest and the imposi-

  tion of a tax on capital would increase the cost of capital and induce more

  capital augmenting innovation rather than labor saving innovation.22

  Non- Pecuniary Externalities

  Innovation may also generate nonpecuniary externalities on agents other

  than the innovator. Classic examples for this are technological externali-

  ties—for example, if an innovation produces public goods or generates or

  alleviates pollution. In markets that deviate from the Arrow- Debreu bench-

  mark, a variety of nonpecuniary eff ects may arise: for example, innovation

  may aff ect quantities demanded, or the probability of buying or selling a

  good or factor, including the probability of being unemployed.

  Some eff ects are such that they can be interpreted either as pecuniary or

  nonpecuniary externalities. For example, product innovations can be inter-

  preted as a price changes—the price of the newly invented good changes

  from infi nity to some positive value—or as a change in the price of the

  consumption services provided by the good. Alternatively, they can also

  be interpreted in a nonpecuniary manner by viewing a product (such as a

  smartphone) as providing a bundle of services to consumers that can only

  be bought in fi xed proportion (e.g., since we cannot separately purchase dif-

  ferent functions of the smartphone). In that view, an innovation represents a

  change in the structure of incomplete markets because it changes the bundle

  of consumption service available from a product. Similarly, changes in job

  quality can be interpreted by viewing each job as a vector of transactions

  that are only available in predetermined bundles, and the innovation changes

  the elements in the bundle that are available. It is well known that changes in

  the degree of market incompleteness for such bundles give rise to externali-

  ties (a specifi c application of Greenwald and Stiglitz 1986).

  14.4 Worker- Replacing Progress and Redistribution

  This section considers a stark form of technological progress that we term

  worker- replacing technological progress. We develop two simple models

  21. As we noted above, there is no fi rst fundamental welfare theorem for innovation, and indeed, there is a presumption that the market is biased toward labor- saving innovation relative to innovations directed toward “saving the planet.” (See Stiglitz 2014b.)

  22. The allocation of resources to capital augmenting technological change depends on the after- tax share of capital. An increase in the relative cost of capital will increase the capital share if the elasticity of substitution is less than unity. Most of the empirical evidence suggests that this is the case.

  AI and Its Implications for Income Distribution and Unemployment 371

  to analyze the two channels generating inequality that we discussed in the

  previous section. In sections 14.4.1 and 14.4.2, we consider the pecuniary

  externalities (redistributions) generated by worker- replacing progress, both

  from a static and a dynamic perspective. In section 14.4.3., we focus on

  the distribution of the surplus accruing to innovators in a model in which

  the surplus is determined by the level of patent protection. Furthermore,

  in section 14.4.4 we discuss the implications of endogenous
factor bias in

  technological progress.

  14.4.1 Static Pecuniary Externalities of Worker- Replacing Progress

  For sections 14.4.1 and 14.4.2, we consider the simple model of worker-

  replacing technological change of Korinek and Stiglitz (2017). We assume

  a production technology that combines capital and labor in a constant-

  returns- to-scale (CRS) function, but where labor consists of the sum of

  human and machine labor. Assuming that human and machine labor enter

  the production function additively means that they are perfect substitutes

  for each other. The details of the baseline model are presented in box 14.1.

  We analyze three questions: What does worker- replacing technological

  change do to wages in the short run and in the long run? And what can

  policy do about it?

  First, we look exclusively at the short run before any of the other factors

  have adjusted:

  Observation 6) Machine Labor and Factor Earnings (in the short run):

  adding a marginal unit of machine labor reduces human wages, but increases

  returns of complementary factors in a zero- sum manner.

  Intuitively, what happens if we add one unit of machine labor is that fi rst,

  that unit will earn its marginal return, but second, there is also a redistribu-

  tion from labor to capital, which now becomes relatively scarcer. The gains

  of capital are exactly the losses of the existing stock of labor.

  The redistribution generated by technological progress can be thought

  of as a pecuniary externality, as we emphasized earlier. The income losses

  of wage earners and the income gains of other factors owners are inef-

  fi cient compared to the fi rst- best benchmark considered in section 14.2.1.

  In the given example, the owners of capital have obtained windfall gains

  but have not done anything to earn these higher return. A compensatory

  transfer from capital owners to workers simply undoes these windfall gains

  and leaves them equally well off as they were before.

  More generally, adding machine labor creates a redistribution away from

  human labor toward complementary factors. This result holds for any CRS

  production function no matter what the complementary factor, for instance,

  whether it is capital or land or unskilled versus skilled labor or entrepre-