The Economics of Artificial Intelligence

Home > Other > The Economics of Artificial Intelligence > Page 62
The Economics of Artificial Intelligence Page 62

by Ajay Agrawal


  neurial rents. Policy can undo these redistributions by taxing windfall gains

  while leaving the price system to work at the margin. The result also holds for

  decreasing- returns- to-scale production functions if we interpret the profi ts

  372 Anton Korinek and Joseph E. Stiglitz

  Box 14.1

  Machine Labor and Factor Earnings

  Assume a constant-returns-to-scale production function that

  produces output Y by combining capital K with labor, consisting

  of the sum of human labor H and machine labor M:

  Y = F( K, H + M).

  In this formulation human labor and machine labor are perfect

  substitutes, so machine technology is what we call worker- replacing.

  In the competitive equilibrium, the wage is determined by the

  marginal product of labor,

  w = F .

  L

  Proposition 1: Machine Labor and Factor Earnings: adding a mar-

  ginal unit of machine labor reduces human wages but increases

  the returns to capital in a zero-sum manner, in addition to increas-

  ing output by the marginal product of labor, which is equal to the

  wage.

  Proof: Using Euler’s Theorem, we rewrite the production func-

  tion:

  ( H + M ) F ( ) + KF ( ) = F( K, H + M ).

  L

  K

  We can now ascertain the eff ect of an additional unit of M:

  F + ( H + M ) F + KF

  = F ,

  L

  LL

  KL

  L

  or, simplifi ed, ( H + M ) F +

  KF

  = 0.

  LL

  KL

  decline in wage bill

  increase in return to K

  Source: Korinek and Stiglitz (2017).

  earned by the owner of the technology as compensation for the implicit

  factor “entrepreneurship,” which takes part in the zero- sum redistribution.

  Let us also emphasize that taxes on previously accumulated factors that

  suddenly earn an unexpected excess return are nondistortionary. This means

  that at least in principle, there is a role for implementing costless redistri-

  bution and generating a Pareto improvement. (In practice, there are some

  natural caveats to this result. For example, it relies on the assumption that

  we can distinguish between previously installed capital that earns windfall

  gains and new capital that would be distorted if it were taxed.)

  AI and Its Implications for Income Distribution and Unemployment 373

  14.4.2 Dynamic Implications of Worker- Replacing Progress

  In the longer run, worker- replacing technological change will lead to sig-

  nifi cant economic change. It implies that the biggest constraint on output—

  the scarcity of labor—is suddenly relaxed. As a result, greater amounts of

  complementary factors, here capital, are accumulated.

  Observation 7) Machine Labor and Abundance of Labor: If not only

  capital, but also labor, is reproducible at suffi

  ciently low cost, then the economy

  will grow exponentially in AK fashion, driven purely by factor accumulation,

  even in the absence of further technological change.

  In Korinek and Stiglitz (2017), we describe the dynamics of this transi-

  tion as machines made by machines get increasingly effi

  cient or, equiva-

  lently, as the cost of producing machines decreases. We identify a singularity

  point at which it becomes cost eff ective for machines to start to fully replace

  human labor.23 In the simplest case, when complementary factors such as

  capital adjust without friction, the human wage may actually be unchanged

  because capital K grows in proportion to eff ective labor ( H + M) so that the marginal productivity of labor and the wage remain unchanged. In other

  words, investment is allocated between conventional machines and human-

  replacing robots in such a way that the return is equal to the intertemporal

  marginal rate of substitution. Under the assumption that workers only care

  about their absolute income, not their relative income, this outcome would

  not be too bad: in absolute terms, even though the human labor share would

  go to zero as an increasing fraction of the labor in the economy is performed

  by machines, workers are no worse off as a result of AI.

  When factors are slow to adjust, the pattern of transition can be com-

  plex, with demand for human labor typically going down temporarily.24 In

  general, the pattern of adjustment depends on how fast the capital stock

  versus the stock of labor adjust. (For example, if the capital stock rises in

  anticipation of an increased supply of machine labor in the future that has

  not yet materialized, then human wages may even go up at intermediate

  stages.)25

  However, the following observation describes that in the long run, workers

  are actually worse off as a result of machine labor if there are nonreproduc-

  ible complementary factors that are in scarce supply, such as land or other

  natural resources.

  23. This singularity captures the important economic aspects of what technologists such as Vernor Vinge (1993) or Ray Kurzweil (2005) call the technological singularity. A similar point is also made in Aghion, Jones, and Jones (2017).

  24. Berg, Buffi

  e, and Zanna (2018) shows that it may actually take decades for the economy’s

  complementary capital stock to adjust after major revolutions in labor- saving technology.

  25. This assumes that capital is “putty- putty,” that is, that capital investments made before AI arrives are equally productive after AI, as would be the case if humans and robots were in fact identical.

  374 Anton Korinek and Joseph E. Stiglitz

  Observation 8) Machine Labor and Return of Scarcity: if there are non-

  reproducible complementary factors, they eventually limit growth; human real

  wages fall, and the owners of nonreproducible factors absorb all the rents.

  Intuitively, as the supply of eff ective labor proliferates due to the intro-

  duction of machine labor, agents in the economy will compete for scarce

  nonreproducible resources like land, driving up their price.

  A similar argument holds for nonreproducible consumption goods: even

  if all factors in the production process are reproducible so that produc-

  tive output in the economy exhibits AK- style growth and workers’ product

  wages remain unchanged, competition for fi xed resources that are part of

  their consumption basket, such as land used for housing, may lead workers

  to eventually be worse off . This may be particularly important in urban set-

  tings where, say, economic activity occurs at the center. Rich rentiers may

  occupy the more desirable locations near the center, with workers having to

  obtain less expensive housing at the periphery, spending more time commut-

  ing. The advent of AI will thus lower their utility.

  However, just as in the earlier case, at the margin, the redistribution from

  workers to nonreproducible factor owners is zero sum. Since taxes on non-

  reproducible factors are by defi nition nondistortionary, there is scope for

  nondistortionary redistribution.

  Observation 9) Nonreproducible Factors and Pareto Improvements: so

  long as nondistortionary taxes on factor rents are feasible, labor- replacing

  innovation can be a Pareto i
mprovement.

  14.4.3 Redistributing the Innovators’ Surplus

  via Changes in Institutions

  If outright redistribution is not feasible or limited, there may be other

  institutional changes that result in market distributions that are more favor-

  able to workers. For example, intervention to steer technological progress

  may act as a second- best device.

  In this section, we provide an example in which a change in intellectual

  property rights—a shortening of the term of patent protection—eff ectively

  redistributes some of the innovators’ surplus to workers (consumers) to

  mitigate the pecuniary externalities on wages that they experience, with the

  ultimate goal that the benefi ts of the innovation are more widely shared.

  If an innovation results in a lower cost of production, then the innovator

  enjoys the benefi ts of the innovation in the form of higher profi ts during the

  life of the patent; but after the expiration of the patent, society enjoys the

  benefi ts in terms of lower prices. The trade- off is that shortening the life of

  the patent may reduce the pace of innovation. But in the spirit of the theory

  of the second- best, there is generally an “optimal” patent life, in which there

  is still some innovation, but in which the well- being of workers is protected.

  Box 14.2

  Intellectual Property Regime and Redistribution

  Consider an economy with a unit mass of workers H = 1, in

  which the capital stock supplied each period K(τ) is a function

  solely of a distortionary capital tax τ, the proceeds of which are

  distributed to workers, and the eff ective stock of machine labor

  M( z) is an increasing function of patent life z.

  A worker’s total income I consists of her wage plus the revenue

  of the capital tax,

  I = w + τ K(τ).

  For any level of M( z), we defi ne τ( M) as the value of the capital tax that keeps workers just as well off as they were before the

  introduction of machine labor.

  Proposition 1. As long as elasticity of capital supply is not too

  large, we can always increase z from z = 0 and compensate work-

  ers by raising the capital tax τ.

  Steady-State Dynamics

  Consider an intertemporal setting in which the growth rate g =

  g ( z, τ) is a function of the length of the patent z and the tax rate τ, by which we now denote the tax rate on innovators. Assume

  that the share of output that is invested is a function of the

  growth rate ( i ( g)) and that the fraction of output not spent on

  investment that is appropriated by the innovator is b( z, τ). In

  steady state, the present discounted value of the income of work-

  ers can be approximated as

  PDV = (1 – i ( g)) [1 – (1 – τ) b( z, τ)]/ ( r – g), where r is the discount rate. If we choose { z, τ} to maximize the

  PDV, in general, the optimum will not be a corner solution in

  which any innovation hurts workers.

  Proposition 2. In general, the optimal { z*, τ*} entails g > 0.

  It is easy to write down suffi

  cient conditions under which Propo-

  sition 2 holds: setting τ* equal to zero, all that we require is that

  | g | is not too large relative to | b |.

  z

  z

  376 Anton Korinek and Joseph E. Stiglitz

  With network externalities the innovator may be able to maintain a domi-

  nant position even after the end of the patent, and may continue to earn the

  surplus from her innovation. With taxes on monopoly profi ts, it should be

  possible to ensure that the innovations are Pareto improving and that even

  human worker- replacing technological change can improve the well- being

  of workers.

  14.4.4 Factor- Biased Technological Change

  So far, we have simply assumed that technological change—the intro-

  duction of AI—is worker replacing. But advances in technology also make

  some machines more productive and others obsolete, aff ecting the (mar-

  ginal) return to traditional capital.26 It is thus useful to think of the world

  as having three groups: capitalists, workers, and innovators. Intellectual

  property rights (and antitrust laws) determine the returns to innovators,

  but the nature of technological change in a competitive market determines

  the division between workers and capitalists.

  A long- standing literature, going back to Kennedy (1964), von Weizacker

  (1966), and Samuelson (1965), describes the endogenous determination of

  the factor bias of technological progress.27 The central result is that as the

  share of labor becomes smaller, the bias shifts toward capital- augmenting

  technological progress. If the world works as these models suggest, this

  should limit the decline in the share of labor (at least in a stable equilib-

  rium) and in inequality.28 As the share of labor decreases, the incentive to

  produce worker- replacing innovation such as AI decreases. But the relevant

  discounted future wage share near the point of singularity—the point where

  it is cost eff ective for machines to fully replace human labor and produce

  more machines all by themselves—may be suffi

  ciently great that there is

  nonetheless an incentive to pass the point of singularity.

  Let us assume that land becomes the binding constraint once human labor

  is fully replaceable by machine labor. In that case, provided the elasticity of

  substitution between land and the other production factors—capital cum

  labor—is less than unity, the share of land increases over time, generating

  the result (analogous to that where labor is the binding constraint in the

  standard literature) that in the long run, all technological progress is land

  augmenting. If the production function is constant returns to scale in land,

  26. As we noted earlier, IA (intelligence assisting) innovation may increase the productivity of humans, and thus increase the demand for humans if the elasticity of substitution is less than unity.

  27. Important contributions were also made by Drandakis and Phelps (1965). More recently, there has been some revival of the literature, with work of Acemoglu (2002), Stiglitz (2006, 2014b) and Acemoglu and Restrepo (2018), among others.

  28. One can describe dynamics with standard wage- setting mechanisms. The system is stable so long as the elasticity of substitution between factors is less than unity (Acemoglu 1998; Stiglitz 2006, 2014b).

  AI and Its Implications for Income Distribution and Unemployment 377

  labor (including machine labor) and traditional capital, then the long- run

  rate of growth is determined by the pace of land- augmenting technological

  change.

  Role of the Service Sector

  Currently, progress in AI focuses on certain sectors of the economy, like

  manufacturing. Partly because of the resulting lower cost of manufacturing,

  and partly because of the shape of preferences, the economy is evolving

  toward a service- sector economy. (If there is diff erential productivity across

  sectors, and the elasticity of demand for the innovation sector is not too

  high, then production factors will move out of that sector into other sec-

  tors. This is even more so if preferences are nonhomothetic, for example,

  demand for food and many manufactured goods having an income elastic-

 
ity less than unity.) Among the key service sectors are education, health,

  the military, and other public services. The value of those services is in

  large part socially determined, that is, by public policies not just a market

  process. If we value those services highly—pay good wages, provide good

  working conditions, and create a suffi

  cient number of jobs—this will limit

  increases in the inequality of market income. Governments typically play

  an important role in these sectors, and their employment policies will thus

  play an important role in the AI transition. Many of these service- sector

  jobs have limited skill requirements. However, higher public- sector wages

  will—through standard equilibrium eff ects—also raise wages in the private

  sector, will improve the bargaining position of workers, and will result in

  such jobs having higher “respect.” All of this will, of course, require tax

  revenues. If the elasticity of entrepreneurial services is low, for example, if

  entrepreneurs are driven partly by nonpecuniary motives, we can impose

  high taxes to fi nance these jobs.

  14.5 Technological

  Unemployment

  Unemployment is one of the most problematic societal implications

  of technological progress—new technology often implies that old jobs

  are destroyed and workers need to fi nd new jobs. Economists, of course,

  understand the “lump- of-labor fallacy”—the false notion that there is a

  fi xed number of jobs, and that automating a given job means that there will

  forever be fewer jobs left in the economy. In a well- functioning economy,

  we generally expect that technological progress creates additional income,

  which in turn can support more jobs.

  However, there are two sound economic reasons for why technological

  unemployment may arise: fi rst, because wages do not adjust for some struc-

  tural reason, as described, for example, by effi

  ciency wage theories, and

  second as a transition phenomenon. The two phenomena may also interact

  378 Anton Korinek and Joseph E. Stiglitz

  in important ways, for example, when effi

  ciency wage considerations slow

  down the transition to a new equilibrium. We discuss the two categories in

  turn in the following subsections.

  The unemployment implications are especially problematic when techno-

  logical progress is labor saving, which—by defi nition—requires that either

 

‹ Prev