The Economics of Artificial Intelligence Read online

  programs measure AI capital.

  The primary diffi

  culty in AI capital measurement is, as mentioned earlier,

  that many of its outputs will be intangible. This issue is exacerbated by

  the extensive use of AI as an input in making other capital, including new

  types of software, as well as human and organizational capital, rather than

  fi nal consumption goods. Much of this other capital, including human

  capital, will, like AI itself, be mostly intangible (Jones and Romer 2010).

  To be more specifi c, eff ective use of AI requires developing data sets,

  building fi rm- specifi c human capital, and implementing new business pro-

  cesses. These all require substantial capital outlays and maintenance. The

  tangible counterparts to these intangible expenditures, including purchases

  of computing resources, servers, and real estate, are easily measured in the

  standard neoclassical growth accounting model (Solow 1957). On the other

  hand, the value of capital goods production for complementary intangible

  investments is diffi

  cult to quantify. Both tangible and intangible capital

  stocks generate a capital service fl ow yield that accrues over time. Real-

  izing these yields requires more than simply renting capital stock. After

  purchasing capital assets, fi rms incur additional adjustment costs (e.g.,

  business process redesigns and installation costs). These adjustment costs

  make capital less fl exible than frictionless rental markets would imply. Much

  of the market value of AI capital specifi cally, and IT capital more gen-

  erally, may be derived from the capitalized short- term quasi- rents earned

  by fi rms that have already reorganized to extract service fl ows from new

  investment.

  Yet while the stock of tangible assets is booked on corporate balance

  sheets, expenditures on the intangible complements and adjustment costs

  to AI investment commonly are not. Without including the production and

  use of intangible AI capital, the usual growth accounting decompositions

  of changes in value added can misattribute AI intangible capital deepening

  Artifi cial Intelligence and the Modern Productivity Paradox 47

  to changes in TFP. As discussed in Hall (2000) and Yang and Brynjolfsson

  (2001), this constitutes an omission of a potentially important component

  of capital goods production in the calculation of fi nal output. Estimates of

  TFP will therefore be inaccurate, though possibly in either direction. In the

  case where the intangible AI capital stock is growing faster than output,

  then TFP growth will be underestimated, while TFP will be overestimated

  if capital stock is growing more slowly than output.

  The intuition for this eff ect is that in any given period t, the output of

  (unmeasured) AI capital stock in period t + 1 is a function the input (unmea-

  sured) existing AI capital stock in period t. When AI stock is growing rapidly, the unmeasured outputs (AI capital stock created) will be greater than the

  unmeasured inputs (AI capital stock used).

  Furthermore, suppose the relevant costs in terms of labor and other

  resources needed to create intangible assets are measured, but the resulting

  increases in intangible assets are not measured as contributions to output. In

  this case, not only will total GDP be undercounted but so will productivity,

  which uses GDP as its numerator. Thus periods of rapid intangible capital

  accumulation may be associated with lower measured productivity growth,

  even if true productivity is increasing.

  With missing capital goods production, measured productivity will only

  refl ect the fact that more capital and labor inputs are used up in producing

  measured output. The inputs used to produce unmeasured capital goods will

  instead resemble lost potential output. For example, a recent report from

  the Brookings Institution estimates that investments in autonomous vehicles

  have topped $80 billion from 2014 to 2017 with little consumer adoption of

  the technology so far.21 This is roughly 0.44 percent of 2016 GDP (spread

  over three years). If all of the capital formation in autonomous vehicles

  was generated by equally costly labor inputs, this would lower estimated

  labor productivity by 0.1 percent per year over the last three years since

  autonomous vehicles have not yet led to any signifi cant increase in mea-

  sured fi nal output. Similarly, according to the AI Index, enrollment in AI

  and ML courses at leading universities has roughly tripled over the past ten

  years, and the number of venture- back AI- related start-ups has more than

  quadrupled. To the extent that they create intangible assets beyond the costs

  of production, GDP will be underestimated.

  Eventually the mismeasured intangible capital goods investments are

  expected to yield a return (i.e., output) by their investors. If and when mea-

  surable output is produced by these hidden assets, another mismeasure-

  ment eff ect leading to overestimation of productivity will kick in. When the

  output share and stock of mismeasured or omitted capital grows, the mea-

  sured output increases produced by that capital will be incorrectly attributed

  to total factor productivity improvements. As the growth rate of invest-

  ment in unmeasured capital goods decreases, the capital service fl ow from

  21. https:// www .brookings .edu / research / gauging - investment - in-self- driving- cars/.

  48 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  unmeasured goods eff ect on TFP can exceed the underestimation error from

  unmeasured capital goods.

  Combining these two eff ects produces a “J- curve” wherein early produc-

  tion of intangible capital leads to underestimation of productivity growth,

  but later returns from the stock of unmeasured capital creates measured

  output growth that might be incorrectly attributed to TFP.

  Formally:

  (1)

  Y + zI = f ( A, K , K , L)

  2

  1

  2

  (2)

  dY + zdI = F dA + F dK + F dL + F dK .

  2

  A

  K

  1

  L

  K

  2

  1

  2

  Output Y and unmeasured capital goods with price z(zI ) are produced

  2

  with production function f. The inputs of f (·) are the total factor productivity A, ordinary capital K , unmeasured capital K , and labor L. Equation (2) 1

  2

  describes the total diff erential of output as a function of the inputs to the

  production function. If the rental price of ordinary capital is r , the rental 1

  price of unmeasured capital is r , and the wage rate is w, we have

  2

  r K

  dK

  wL

  dL

  (3)

  ˆ

  S = dY

  1

  1

  1

  Y

  Y

  K

  Y

  L

  1

  and

  r K

  dK

  wL

  dL

  r K

  dK

  dI

  (4)

  S* = dY

  1

  1

  1
/>
  2

  2

  2

  + zI 2

  2

  ,

  Y

  Y

  K

  Y

  L

  Y

  K

  Y

  I

  1

  2

  2

  where ˆ

  S is the familiar Solow residual as measured and S∗ is the correct

  Solow residual accounting for mismeasured capital investments and stock.

  The mismeasurement is then

  K

  dK

  zI

  dI

  K

  zI

  (5)

  ˆ

  S

  S* = r 2 2

  2

  2

  2

  = r 2 2 g

  2

  g .

  Y

  K

  Y

  I

  Y

  K

  I

  2

  Y

  2

  2

  2

  The right side of the equation describes a hidden capital eff ect and a hidden

  investment eff ect. When the growth rate of new investment in unmeasured

  capital multiplied by its share of output is larger (smaller) than the growth

  rate of the stock of unmeasured capital multiplied by its share of output,

  the estimated Solow residual will underestimate (overestimate) the rate of

  productivity growth. Initially, new types of capital will have a high marginal

  product. Firms will accumulate that capital until its marginal rate of return

  is equal to the rate of return of other capital. As capital accumulates, the

  growth rate of net investment in the unmeasured capital will turn negative,

  causing a greater overestimate TFP. In steady state, neither net investment’s

  share of output nor the net stock of unmeasured capital grows and the pro-

  ductivity mismeasurement is zero. Figure 1.9 provides an illustration.22

  22. The price of new investment ( z) and rental price of capital ( r) are 0.3 and 0.12, respectively, in this toy economy. Other values used to create the fi gure are included in the appendix.

  Artifi cial Intelligence and the Modern Productivity Paradox 49

  Fig. 1.9 The mismeasurement J- curve for an economy accumulating a new kind

  of capital

  Looking forward, these problems may be particularly stark for AI capital,

  because its accumulation will almost surely outstrip the pace of ordinary

  capital accumulation in the short run. AI capital is a new category of

  capital—new in economic statistics, certainly, but we would argue practi-

  cally so as well.

  This also means that capital quantity indexes that are computed from

  within- type capital growth might have problems benchmarking size and

  eff ect of AI early on. National statistics agencies do not really focus on mea-

  suring capital types that are not already ubiquitous. New capital categories

  will tend to either be rolled into existing types, possibly with lower inferred

  marginal products (leading to an understatement of the productive eff ect

  of the new capital), or missed altogether. This problem is akin to the new

  goods problem in price indexes.

  A related issue—once AI is measured separately—is how closely its units

  of measurement will capture AI’s marginal product relative to other capital

  stock. That is, if a dollar of AI stock has a marginal product that is twice

  as high as the modal unit of non- AI capital in the economy, will the quan-

  tity indexes of AI refl ect this? This requires measured relative prices of AI

  and non- AI capital to capture diff erences in marginal product. Measuring

  levels correctly is less important than measuring accurate proportional dif-

  ferences (whether intertemporally or in the cross section) correctly. What is

  needed in the end is that a unit of AI capital twice as productive as another

  should be twice as large in the capital stock.

  It is worth noting that these are all classic problems in capital measure-

  ment and not new to AI. Perhaps these problems will be systematically worse

  for AI, but this is not obvious ex ante. What it does mean is that econo-

  50 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  mists and national statistical agencies at least have experience in, if not

  quite a full solution for, dealing with these sorts of limitations. That said,

  some measurement issues are likely to be especially prevalent for AI. For

  instance, a substantial part of the value of AI output may be fi rm- specifi c.

  Imagine a program that fi gures out individual consumers’ product prefer-

  ences or price elasticities and matches products and pricing to predictions.

  This has diff erent value to diff erent companies depending on their customer

  bases and product selection, and knowledge may not be transferrable across

  fi rms. The value also depends on companies’ abilities to implement price

  discrimination. Such limits could come from characteristics of a company’s

  market, like resale opportunities, which are not always under fi rms’ control,

  or from the existence in the fi rm of complementary implementation assets

  and/or abilities. Likewise, each fi rm will likely have a diff erent skill mix that

  it seeks in its employees, unique needs in its production process, and a par-

  ticular set of supply constraints. In such cases, fi rm- specifi c data sets and

  applications of those data will diff erentiate the machine- learning capabili-

  ties of one fi rm from another (Brynjolfsson and McAfee 2017).

  1.11 Conclusion

  There are plenty of both optimists and pessimists about technology and

  growth. The optimists tend to be technologists and venture capitalists, and

  many are clustered in technology hubs. The pessimists tend to be econo-

  mists, sociologists, statisticians, and government offi

  cials. Many of them are

  clustered in major state and national capitals. There is much less interaction

  between the two groups than within them, and it often seems as though they

  are talking past each other. In this chapter, we argue that in an important

  sense, they are.

  When we talk with the optimists, we are convinced that the recent break-

  throughs in AI and machine learning are real and signifi cant. We also would

  argue that they form the core of a new, economically important potential

  GPT. When we speak with the pessimists, we are convinced that productiv-

  ity growth has slowed down recently and what gains there have been are

  unevenly distributed, leaving many people with stagnating incomes, declin-

  ing metrics of health and well- being, and good cause for concern. People

  are uncertain about the future, and many of the industrial titans that once

  dominated the employment and market value leaderboard have fallen on

  harder times.

  These two stories are not contradictory. In fact, in many ways they are

  consistent and symptomatic of an economy in transition. Our analysis sug-

  gests that while the recent past has been diffi

  cult, it is not destiny. Although

  it is always dangerous to make predictions, and we are humble about our

  ability to foretell the future, our reading of the evidence does provide some

  cause for optimism. The
breakthroughs of AI technologies already demon-

  Artifi cial Intelligence and the Modern Productivity Paradox 51

  strated are not yet aff ecting much of the economy, but they portend big-

  ger eff ects as they diff use. More important, they enable complementary

  innovations that could multiply their impact. Both the AI investments and

  the comple mentary changes are costly, hard to measure, and take time to

  implement, and this can, at least initially, depress productivity as it is cur-

  rently measured. Entrepreneurs, managers, and end- users will fi nd powerful

  new applications for machines that can now learn how to recognize objects,

  understand human language, speak, make accurate predictions, solve prob-

  lems, and interact with the world with increasing dexterity and mobility.

  Further advances in the core technologies of machine learning would

  likely yield substantial benefi ts. However, our perspective suggests that an

  underrated area of research involves the complements to the new AI tech-

  nologies, not only in areas of human capital and skills, but also new processes

  and business models. The intangible assets associated with the last wave of

  computerization were about ten times as large as the direct investments in

  computer hardware itself. We think it is plausible that AI- associated intan-

  gibles could be of a comparable or greater magnitude. Given the big changes

  in coordination and production possibilities made possible by AI, the ways

  that we organized work and education in the past are unlikely to remain

  optimal in the future.

  Relatedly, we need to update our economic measurement tool kits. As

  AI and its complements more rapidly add to our (intangible) capital stock,

  traditional metrics like GDP and productivity can become more diffi

  cult to

  measure and interpret. Successful companies do not need large investments

  in factories or even computer hardware, but they do have intangible assets

  that are costly to replicate. The large market values associated with compa-

  nies developing and/or implementing AI suggest that investors believe there

  is real value in those companies. In the case that claims on the assets of the

  fi rm are publicly traded and markets are effi

  cient, the fi nancial market will

  properly value the fi rm as the present value of its risk- adjusted discounted

  cash fl ows. This can provide an estimate of the value of both the tangible

  and intangible assets owned by the fi rm. What’s more, the eff ects on living

  standards may be even larger than the benefi ts that investors hope to cap-