Book Read Free

The Economics of Artificial Intelligence

Page 45

by Ajay Agrawal

potential singularity eff ects might aff ect growth and convergence.

  A fi rst idea is that new AI technologies might allow imitation/ learning of

  frontier technologies to become automated. That is, machines would fi gure

  out in no time how to imitate frontier technologies. Then a main source

  of divergence might become credit constraints, to the extent that those

  might prevent poorer countries or regions from acquiring superintelligent

  machines whereas developed economies could aff ord such machines. Thus

  one could imagine a world in which advanced countries concentrate all their

  research eff ort on developing new product lines (i.e., on frontier innovation)

  whereas poorer countries would devote a positive and increasing fraction of

  their research labor on learning about the new frontier technologies as they

  cannot aff ord the corresponding AI devices. Overall, one would expect an

  increasing degree of divergence worldwide.

  A second conjecture is that, anticipating the eff ect of AI on the scope and

  speed of imitation, potential innovators may become reluctant to patent

  their inventions, fearing that the disclosure of new knowledge in the patent

  would lead to straight imitation. Trade secrets may then become the norm,

  instead of patenting. Or alternatively innovations would become like what

  fi nancial innovations are today, that is, knowledge creation with huge net-

  work eff ects and with very little scope for patenting.

  Finally, with imitation and learning being performed mainly by super-

  machines in developed economies, then research labor would become

  (almost) entirely devoted to product innovation, increasing product variety

  or inventing new products (new product lines) to replace existing products.

  Then, more than ever, the decreasing returns to digging deeper into an ex-

  isting line of product would be off set by the increased potential for discov-

  ering new product lines. Overall, ideas might end up being easier to fi nd,

  if only because of the singularity eff ect of AI on recombinant idea- based

  growth.

  262 Philippe Aghion, Benjamin F. Jones, and Charles I. Jones 9.5 Artifi cial Intelligence, Firms, and Economic Growth

  To this point, we have linked artifi cial intelligence to economic growth

  emphasizing features of the production functions of goods and ideas. How-

  ever, the advance of artifi cial intelligence and its macroeconomic eff ects

  will depend on the potentially rich behavior of fi rms. We have introduced

  one such view already in the prior section, where considerations of crea-

  tive destruction provide an incentive- oriented mechanism that may be an

  important obstacle to singularities. In this section, we consider fi rms’ incen-

  tives and behavior more generally to further outline the AI research agenda.

  We examine potentially fi rst- order issues that emerge when introducing

  market structure, sectoral diff erences, and organizational considerations

  within fi rms.

  9.5.1 Market

  Structure

  Existing work on competition and innovation- led growth points to the

  existence of two counteracting eff ects: on the one hand, more intense prod-

  uct market competition (or imitation threat) induces neck- and- neck fi rms

  at the technological frontier to innovate in order to escape competition; on

  the other hand, more intense competition tends to discourage fi rms behind

  the current technology frontier to innovate and thereby catch-up with fron-

  tier fi rms. Which of these two eff ects dominates, in turn, depends upon

  the degree of competition in the economy, and/or upon how advanced the

  economy is. While the escape competition eff ect tends to dominate at low

  initial levels of competition and in more advanced economies, the discour-

  agement eff ect may dominate for higher levels of competition or in less

  advanced economies.18

  Can AI aff ect innovation and growth through potential eff ects it might

  have on product market competition? A fi rst potential channel is that AI

  may facilitate the imitation of existing products and technologies. Here we

  particularly have in mind the idea that AI might facilitate reverse engineer-

  ing, and thereby facilitate the imitation of leading products and technolo-

  gies. If we follow the inverted- U logic of Aghion et al. (2005), in sectors

  with initially low levels of imitation, some AI- induced reverse engineering

  might stimulate innovation by virtue of the escape- competition eff ect. But

  too high (or too immediate) an imitation threat will end up discourag-

  ing innovation as potential innovators will face excessive expropriation.

  A related impli cation of AI is that its introduction may speed up the pro-

  cess by which each individual sector becomes congested over time. This in

  turn may translate into faster decreasing returns to innovating within any

  existing sector (see Bloom et al. 2014), but by the same token it may induce

  potential innovators to devote more resources to inventing new lines in

  18. For example, see Aghion and Howitt (1992) and Aghion et al. (2005).

  Artifi cial Intelligence and Economic Growth 263

  order to escape competition and imitation within current lines. The overall

  eff ect on aggregate growth will in turn depend upon the relative contribu-

  tions of within- sector secondary innovation and fundamental innovation

  aimed at creating new product lines (see Aghion and Howitt 1996) to the

  overall growth process.

  Another channel whereby AI and the digital revolution may aff ect inno-

  vation and growth through aff ecting the degree of product market compe-

  tition is in relation to the development of platforms or networks. A main

  objective of platform owners is to maximize the number of participants to

  the platform on both sides of the corresponding two- sided markets. For ex-

  ample, Google enjoys a monopoly position as a search platform, Facebook

  enjoys a similar position as a social network with more than 1.7 billion

  users worldwide each month, and so does Booking .com for hotel reserva-

  tions (more than 75 percent of hotel clients resort to this network). And

  the same goes for Uber in the area of individual transportation, Airbnb for

  apartment renting, and so on. The development of networks may in turn

  aff ect competition in at least two ways. First, data access may act as an entry

  barrier for creating new competing networks, although it did not prevent

  Facebook from developing a new network after Google. More important,

  networks can take advantage of their monopoly positions to impose large

  fees on market participants (and they do), which may discourage innovation

  by these participants, whether they are fi rms or self- employed individuals.

  In the end, whether escape competition or discouragement eff ects domi-

  nate will depend upon the type of sector (frontier/ neck- and- neck or older/

  lagging), the extent to which AI facilitates reverse engineering and imita-

  tion, and upon competition and/or regulatory policies aimed at protecting

  intellectual property rights while lowering entry barriers. Recent empirical

  work (e.g., see Aghion, Howitt, and Prantl 2015) points at patent protection

  and competition policy being
complementary in inducing innovation and

  productivity growth. It would be interesting to explore how AI aff ects this

  complementarity between the two policies.

  9.5.2 Sectoral Reallocation

  A recent paper by Baslandze (2016) argues that the information tech-

  nology (IT) revolution has produced a major knowledge diff usion eff ect,

  which in turn has induced a major sectoral reallocation from sectors that

  do not rely much on technological externalities from other fi elds or sectors

  (e.g., textile industries) to sectors that rely more heavily on technological

  externalities from other sectors. Her argument, which we believe applies

  to AI, rests on the following two counteracting eff ects of IT on innovation

  incentives: on the one hand, fi rms can more easily learn from each other and

  therefore benefi t more from knowledge diff usion from other fi rms and sec-

  tors; on the other hand, the improved access to knowledge from other fi rms

  and sectors induced by IT (or AI) increases the scope for business stealing.

  264 Philippe Aghion, Benjamin F. Jones, and Charles I. Jones In high- tech sectors where fi rms benefi t more from external knowledge, the

  former eff ect—knowledge diff usion—will dominate whereas in sectors that

  do not rely much on external knowledge the latter eff ect—competition or

  business stealing—will tend to dominate. Indeed in more knowledge depen-

  dent sectors fi rms see both their productive and their innovative capabilities

  increase to a larger extent than the capabilities of fi rms in sectors that rely

  less on knowledge from other sectors.

  It then immediately follows that the diff usion of IT—and AI for our

  purpose—should lead to an expansion of sectors that rely more on exter-

  nal knowledge (in which the knowledge diff usion eff ect dominates) at the

  expense of the more traditional (and more self- contained) sectors where

  fi rms do not rely as much on external knowledge.

  Thus, in addition to its direct eff ects on fi rms’ innovation and production

  capabilities, the introduction of IT and AI involve a knowledge diff usion

  eff ect that is augmented by a sectoral reallocation eff ect at the benefi t of

  high- tech sectors that rely more on knowledge externalities from other fi elds

  and sectors. The positive knowledge diff usion eff ect is partly counteracted by

  the negative business- stealing eff ect (Baslandze shows that the latter eff ect

  has been large in the United States and that without it the IT revolution

  would have yet induced a much higher acceleration in productivity growth

  for the whole US economy).

  Based on her analysis, Baslandze (2016) responds to Gordon (2012) with

  the argument that Gordon only took into account the direct eff ect of IT

  and not its indirect knowledge diff usion and sectoral reallocation eff ects on

  aggregate productivity growth.

  We believe that the same points can be made with respect to AI instead

  of IT, and one could try and reproduce Baslandze’s calibration exercise to

  assess the relative importance of the direct and indirect eff ects of AI, to

  decompose the indirect eff ect of AI into its positive knowledge diff usion

  eff ect and its potentially negative competition eff ect, and to assess the extent

  to which AI aff ects overall productivity growth through its eff ects on sectoral

  reallocation.

  9.5.3 Organization

  How should we expect fi rms to adapt their internal organization, the skill

  composition of their workforce and their wage policies to the introduction

  of AI? In his recent book, Economics for the Common Good, Tirole (2017)

  spells out what one may consider to be “common wisdom” expectations

  on fi rms and AI. Namely, introducing AI should: (a) increase the wage gap

  between skilled and unskilled labor, as the latter is presumably more sub-

  stitutable to AI than the former; (b) the introduction of AI allows fi rms

  to automate and dispense with middle men performing monitoring tasks

  (in other words, fi rms should become fl atter, that is, with higher spans of

  control); (c) should encourage self- employment by making it easier for indi-

  Artifi cial Intelligence and Economic Growth 265

  viduals to build their reputation. Let us revisit these various points in more

  detail. AI, skills, and wage premia: on AI and the increased gap between

  skilled and unskilled wage, the prediction brings us back to Krusell et al.

  (2000) based on an aggregate production function in which physical equip-

  ment is more substitutable to unskilled labor than to skilled labor, these

  authors argued that the observed acceleration in the decline of the relative

  price of production equipment goods since the mid- 1970s could account for

  most of the variation in the college premium over the past twenty- fi ve years.

  In other words, the rise in the college premium could largely be attributed

  to an increase in the rate of (capital- embodied) skill- biased technical pro-

  gress. And, presumably, AI is an extreme form of capital- embodied, skill-

  biased technical change, as robots substitute for unskilled labor but require

  skilled labor to be installed and exploited. However, recent work by Aghion

  et al. (2017) suggests that while the prediction of a premium to skills may

  hold at the macroeconomic level, it perhaps misses important aspects of

  fi rms’ internal organization and that the organization itself may evolve as a

  result of introducing AI. More specifi cally, Aghion et al. (2017) use matched

  employer- employee data from the United Kingdom, which they augment

  with information on R&D expenditures, to analyze the relationship between

  innovativeness and average wage income across fi rms.

  A fi rst, not surprising, fi nding is that more R&D-intensive fi rms pay

  higher wages on average and employ a higher fraction of high- occupation

  workers than less R&D-intensive fi rms (see fi gure 9.4).

  This, in turn, is perfectly in line with the above prediction (a) but also with

  prediction (b) as it suggests that more innovative (or more “frontier” ) fi rms

  rely more on outsourcing for low- occupation tasks. However, a more sur-

  prising fi nding in Aghion et al. (2017) is that lower- skill (lower occupation)

  workers benefi t more from working in more R&D-intensive fi rms (relative

  to working in a fi rm that does no R&D) than higher- skill workers. This fi nd-

  ing is summarized by fi gure 9.5. In that fi gure, we fi rst see that higher- skill

  workers earn more than lower- skill workers in any fi rm no matter how R&D

  intensive that fi rm is (the high- skill wage curve always lies strictly above the

  middle- skill curve, which itself always lies above the lower- skill curve). But,

  more interestingly, the lower- skill curve is steeper than the middle- skill and

  higher- skill curve. But the slope of each of these curves precisely refl ects

  the premium for workers with the corresponding skill level to working in a

  more innovative fi rm.

  Similarly, we should expect more AI- intensive fi rms to: (a) employ a

  higher fraction of (more highly paid) high- skill workers, (b) outsource an

  increasing fraction of low- occupation tasks, and (c) give a higher premium

  to those low-
occupation workers they keep within the fi rm (unless we take

  the extreme view that all the functions to be performed by low- occupation

  workers could be performed by robots).

  To rationalize the above fi ndings and these latter predictions, let us fol-

  Fig. 9.4 Log hourly wage and R&D intensity

  Source: Aghion et al. (2017).

  Note: This fi gure plots the logarithm of total hourly income against the logarithm of total R&D expenditures (intramural + extramural) per employee (R&D intensity).

  Fig. 9.5 Log hourly wage and R&D intensity

  Source: Aghion et al. (2017).

  Note: This fi gure plots the logarithm of total hourly income against the logarithm of total R&D expenditures (intramural + extramural) per employee (R&D intensity) for diff erent skill groups.

  Artifi cial Intelligence and Economic Growth 267

  low Aghion et al. (2017) who propose a model in which more innovative

  fi rms display a higher degree of complementarity between low- skill workers

  and the other production factors (capital and high- skill labor) within the

  fi rm. Another feature of their model is that high- occupation employees’

  skills are less fi rm- specifi c than low- skill workers: namely, if the fi rm was

  to replace a high- skill worker by another high- skill worker, the downside

  risk would be limited by the fact that higher- skill employees are typically

  more educated employees, whose market value is largely determined by their

  education and accumulated reputation, whereas low- occupation em ployees’

  quality is more fi rm- specifi c. This model is meant to capture the idea that

  low- occupation workers can have a potentially more damaging eff ect on

  the fi rm’s value if the fi rm is more innovative (or more AI intensive for our

  purpose).

  In particular, an important diff erence with the common wisdom, is that

  here innovativeness (or AI intensity) impacts on the organizational form of

  the fi rm and in particular on complementarity or substitutability between

  workers with diff erent skill levels within the fi rm, whereas the common wis-

  dom view takes this complementarity or substitutability as given. Think

  of a low- occupation employee (e.g., an assistant) who shows outstanding

  ability, initiative, and trustworthiness. That employee performs a set of

  tasks for which it might be diffi

 

‹ Prev