The Economics of Artificial Intelligence

Home > Other > The Economics of Artificial Intelligence > Page 38
The Economics of Artificial Intelligence Page 38

by Ajay Agrawal


  already discussed in section 8.2, the history of technology of the last two cen-

  turies is full of examples of automation, ranging from weaving and spinning

  machines to the mechanization of agriculture, as discussed in the previous

  section. Even with capital accumulation and the deepening of automation,

  if there were no other counteracting force, we would see the share of labor

  in national income declining steadily. Our conceptual framework highlights

  a major force preventing such a decline—the creation of new tasks in which

  labor has a comparative advantage.

  This can be seen by putting together equations (7) and (10), which yields

  R

  W

  (11)

  d ln W = ln

  ln

  dN

  ( N

  1)

  ( N )

  M

  L

  W

  R

  + ln

  ln

  dI +

  1

  ( dN dI ),

  ( I )

  ( I )

  N

  I

  L

  M

  and also from equation (6),

  ds = dN

  dI.

  L

  For the labor share to remain stable and for wages to increase in tandem

  with productivity, as has been the case historically, we need I—capturing

  the extensive margin of automation—to grow by the same amount as the

  range of new tasks, N. When that happens, equilibrium wages grow propor-

  tionately with productivity, and the labor share, s , remains constant, as can

  L

  be seen from the fact that the fi rst line of equation (11) is in this case equal

  to the increase in productivity or gross domestic product (GDP) per worker.

  Indeed, rewriting equation (11) imposing dN = dI, we have

  ( N )

  ( I )

  d ln W = ln

  L

  ln

  L

  dI > 0,

  ( N

  1)

  ( I )

  M

  M

  which is strictly positive because of assumption (A1).

  8.4.6 A False Dichotomy: Recap

  With our conceptual framework exposited in a more systematic manner,

  we can now briefl y revisit the false dichotomy highlighted in the introduc-

  tion. Our analysis (in particular equation [7]) highlights that there is always

  a negative displacement eff ect on labor resulting from automation. Equa-

  tion (11) reiterates that there is no presumption that this displacement eff ect

  could not reduce overall demand for labor.

  However, several countervailing eff ects imply that a negative impact from

  automation on labor demand is not a forgone conclusion. Most important,

  the productivity eff ect could outweigh the displacement eff ect, leading to an

  expansion in labor demand and equilibrium wages from automation. The

  Artifi cial Intelligence, Automation, and Work 219

  presence of the productivity eff ect as counterweight to the displacement

  created by automation highlights an important conceptual issue, however.

  In contrast to the emphasis in the popular discussions it is not the brilliant,

  superproductive automation technologies that threaten labor, but the “so-

  so” ones that create the displacement eff ect as they replace labor in tasks

  that it previously performed, but do not engender the countervailing pro-

  ductivity eff ect.

  The productivity eff ect is supplemented by the capital accumulation

  that automation sets in motion and the deepening of automation, which

  increases the productivity of machines in tasks that have already been auto-

  mated. But even with these countervailing eff ects, equation (9) shows that

  automation will always reduce the share of labor in national income. All the

  same, this does not signal the demise of labor either, because the creation

  of new tasks in which labor has a comparative advantage could counterbal-

  ance automation, which is our interpretation of why the demand for labor

  has kept up with productivity growth in the past despite several rapid waves

  of automation.

  Our framework suggests that the biggest shortcoming of the alarmist

  and the optimist views is their failure to recognize that the future of labor

  depends on the balance between automation and the creation of new tasks.

  Automation will often lead to a healthy growth of labor demand and wages

  if it is accompanied with a commensurate increase in the set of tasks in

  which labor has a comparative advantage—a feature that alarmists seem to

  ignore. Even though there are good economic reasons for why the economy

  will create new tasks, this is neither a forgone conclusion nor something

  we can always count on—as the optimists seem to assume. Artifi cial intel-

  ligence and robotics could be permanently altering this balance, causing

  automation to pace ahead of the creation of new tasks with negative conse-

  quences for labor, at the very least in regard to the share of labor in national

  income.

  8.4.7 Generalizations

  Many of the features adopted in the previous subsection are expositional

  simplifi cations. In particular, the aggregate production function (1) can be

  taken to be any constant elasticity of substitution aggregate. One impli-

  cation of this would be that aggregate output in equation (3) would be a

  constant elasticity aggregate itself. This does not aff ect any of our main

  conclusions, including the negative impact of automation on the labor share

  (see Acemoglu and Restrepo 2016).10

  We also do not need assumption (A1) for any of the results. If the second

  10. Recent work by Aghion, Jones, and Jones (2017) points out, however, that if the elasticity of substitution between tasks is less than one and there is an exogenous and high saving rate, the labor share might asymptote to a positive value even with continuously ongoing auto mation.

  220 Daron Acemoglu and Pascual Restrepo

  inequality in this assumption does not hold, changes in automation tech-

  nology have no impact on the equilibrium because it is not cost eff ective to

  adopt all available automation technologies (for this reason, in the general

  case, Acemoglu and Restrepo [2016] distinguish technologically automated

  tasks from equilibrium automation). Given our focus here, there is no loss

  of generality in making this assumption.

  A fi nal feature that is worth commenting on is the fact that in the aggregate

  production function (1), the limits of integration are N – 1 and N, ensuring that the total measure of tasks is one. This is useful for several reasons. First,

  when the introduction of new tasks expands the total measure of tasks, it

  becomes more challenging to obtain a balanced growth path (see Acemoglu

  and Restrepo 2016). Second, in this case some minor modifi cations are nec-

  essary so that an expansion in the total measure of tasks leads to productiv-

  ity improvements. In particular, consider the general case where the elastic-

  ity of substitution between tasks is not necessarily equal to one. If it is

  greater than one, an increase in N leads to higher productivity, but not nec-

  essarily when it is less than or equal to one. In this latter case, we then need

  to introduce direct productiv
ity gains from task variety. For example, in

  the present case where the elasticity of substitution between tasks is equal

  to one, we could modify (1) to ln Y = (1 / N)

  N

  0 ln[ N 1+ y( i )], where ≥ 0

  represents these productivity gains from task variety and ensures that the

  qualitative results explicit here continue to apply.

  8.4.8 Employment

  and

  Unemployment

  An additional generalization concerns the endogenous adjustment of

  employment in the face of new automation technologies. We have so far

  taken labor to be supplied inelastically for simplicity. There are two ways in

  which the level of employment responds to the arrival of new technologies.

  The fi rst is via a standard labor supply margin. Acemoglu and Restrepo

  (2016) show that the endogenous adjustment of labor supply, including

  income eff ects and the substitution of consumption and leisure, links the

  level of employment to the share of labor in national income.

  The second possibility is through labor market frictions, for example, as

  in Acemoglu and Restrepo (2018a). Under appropriate assumptions, the

  endogenous level of employment in this case is also a function of the share

  of labor in national income. Though both models with and without labor

  market frictions endogenize employment as a function of the labor share,

  their normative implications are potentially diff erent, as we discuss below.

  For now, however, the more important implication of such extensions

  is to link the level of employment (or unemployment) to labor demand.

  Automation, when it reduces labor demand, will also reduce the level of

  employment (or increase the level of unemployment). Moreover, because the

  supply of labor depends on the labor share, in our framework automation

  results in a reduction in employment (or an increase in unemployment). As

  such, our analysis so far also sheds light on (and clarifi es the conditions for)

  Artifi cial Intelligence, Automation, and Work 221

  the claims that new automation technologies will reduce employment. It also

  highlights, however, that the fact that automation has been ongoing does not

  condemn the economy to a declining path of employment. If automation is

  met by equivalent changes in the creation of new tasks, the share of labor in

  national income can remain stable and ensure a stable level of employment

  (or unemployment) in the economy.

  8.5 Constraints and Ineffi

  ciencies

  Even in the presence of the countervailing forces limiting the displace-

  ment eff ect from automation, there are potential ineffi

  ciencies and con-

  straints limiting the smooth adjustment of the labor market and hindering

  the productivity gains from new technologies.

  Here we focus on how the mismatch between skills and technologies not

  only increases inequality, but also hinders the productivity gains from auto-

  mation and new tasks. We then explore the possibility that, concurrent with

  rapid automation, we are experiencing a slowdown in the creation of new

  tasks, which could result in slow productivity growth. Finally, we examine

  how a range of factors leads to excessive automation, which not only creates

  ineffi

  ciency but also hinders productivity.

  8.5.1 Mismatch of Technologies and Skills

  The emphasis on the creation of new tasks counterbalancing the potential

  negative eff ects of automation on the labor share and the demand for labor

  ignores an important caveat and constraint: the potential mismatch between

  the requirements of new technologies (tasks) and the skills of the workforce.

  To the extent that new tasks require skilled employees or even new skills to

  be acquired, the adjustment may be much slower than our analysis so far

  suggests.

  To illustrate these ideas in the simplest possible fashion, we follow Acemo-

  glu and Restrepo (2016) and assume that there are two types of workers,

  low- skill with supply L and high- skill with supply H , both of them supplied inelastically. We also assume that low- skill workers can only perform

  tasks below a threshold S ∈ ( I, N), while high- skill workers can perform all tasks. For simplicity, we assume that the productivity of both low- skill

  and high- skill workers in the tasks that they can perform is still given by

  ( x).11 Low- skill workers earn a wage W and high- skill workers earn a L

  L

  wage W .

  H

  11. We can also introduce diff erential comparative advantages and also an absolute productivity advantage for high- skill workers, though we choose not to do so to increase transparency (see Acemoglu and Restrepo 2016). The more restrictive assumption here is that automation happens at the bottom of the range of tasks. In general, automation could take place in the middle range, and its impact would depend on whether automated tasks are competing pre-dominantly against low- skill or high- skill workers (see Acemoglu and Autor 2011; Acemoglu and Restrepo 2018b).

  222 Daron Acemoglu and Pascual Restrepo

  In this simple extension of the framework presented so far, the threshold

  S can be considered as an inverse measure of the mismatch between new

  technologies and skills. A greater value of S implies that there are plenty

  of additional tasks for low- skill workers, while a low value of S implies the presence of only a few tasks left that low- skill workers can perform.

  Assuming that in equilibrium W > W ,12 which implies that low- skill H

  L

  workers will perform all tasks in the range ( I, S), equilibrium wages satisfy W = Y ( N

  S ) and W = Y ( S

  I ).

  H

  H

  L

  L

  Thus, the impact of automation on inequality—defi ned here as the wage

  premium between high- and low- skill workers—is given by

  d ln W / W

  H

  L =

  1

  > 0.

  dI

  S

  I

  This equation shows that automation increases inequality. This is not sur-

  prising, since the tasks that are automated are precisely those performed by

  low- skill workers. But in addition, it also demonstrates that the impact of

  automation on inequality becomes worse when there is a severe skill mis-

  match—the threshold S is close to I. In this case, displaced workers will be squeezed into a very small range of tasks, and hence, each of these tasks will

  receive a large number of workers and will experience a substantial drop in

  price, which translates into a sharp decline in the wage of low- skill workers.

  In contrast, when S is large, displaced workers can spread across a larger set

  of tasks without depressing their wage as much.

  A severe mismatch also aff ects the productivity gains from automation.

  In particular, we have

  d ln( Y / L)

  W

  R

  = ln

  L

  ln

  > 0.

  dI

  ( I )

  ( I )

  L

  M

  This equation shows that the productivity gains from automation depend

  positively on W / R: it is precisely when displaced workers have a high oppor-L

  tunity cost that automation raises productivity. Using the fact that R =


  ( Y/ K)( I – N + 1), we obtain

  W

  K

  L =

  S

  I

  .

  R

  I

  N + 1 L

  A worse mismatch (a lower S) reduces the opportunity cost of displaced

  workers further, and via this channel, it makes automation less profi table.

  This is because a severe mismatch impedes reallocation, reducing the pro-

  ductivity gains of freeing workers from automated tasks.

  12. This is equivalent to [( N – S)/ ( S – I)] > ( H/ L), so that high- skill workers are scarce relative to the range of tasks that only they can produce.

  Artifi cial Intelligence, Automation, and Work 223

  Equally important are the implications of a skill mismatch for the pro-

  ductivity gains from new tasks. Namely,

  d ln( Y / L)

  R

  W

  = ln

  ln

  H

  > 0,

  dN

  ( N

  I )

  ( N )

  M

  H

  which depends negatively on W / R: it is precisely when high- skill workers H

  have a relatively high wage that the gains from new tasks will be limited. With

  similar arguments to before, we also have

  W

  K

  H

  = N S

  ,

  R

  I

  N + 1 L

  which implies that in the presence of a worse mismatch (a lower S), the

  productivity gains from new tasks will be limited. This is because new tasks

  require high- skill workers who are scarce and expensive when S is low.

  An important implication of this analysis is that to limit increasing

  inequality and to best deploy new tasks and harness the benefi ts of auto-

  mation, society may need to simultaneously increase the supply of skills. A

  balanced growth process requires not only automation and the creation of

  new tasks to go hand- in-hand, but also the supply of high- skill workers to

  grow in tandem with these technological trends.

  8.5.2 Automation at the Expense of New Tasks

  As discussed in section 8.2, a puzzling aspect of recent macroeconomic

  developments has been the lack of robust productivity growth despite the

  bewildering array of new technologies. Our conceptual framework provides

  three novel (and at least to us, more compelling) reasons for slow produc-

  tivity growth. The fi rst was the skill mismatch discussed in the previous

 

‹ Prev