The Economics of Artificial Intelligence Read online

  in capital’s current or expected marginal product.

  These facts have been read by some as reasons for pessimism about the

  ability of new technologies like AI to greatly aff ect productivity and income.

  Gordon (2014, 2015) argues that productivity growth has been in long- run

  decline, with the IT- driven acceleration of 1995 to 2004 being a one- off

  aberration. While not claiming technological progress will be nil in the com-

  ing decades, Gordon essentially argues that we have been experiencing the

  new, low- growth normal and should expect to continue to do so going for-

  ward. Cowen (2011) similarly off ers multiple reasons why innovation may

  be slow, at least for the foreseeable future. Bloom et al. (2017) document

  28 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  that in many fi elds of technological progress research productivity has been

  falling, while Nordhaus (2015) fi nds that the hypothesis of an acceleration

  of technology- driven growth fails a variety of tests.

  This pessimistic view of future technological progress has entered into

  long- range policy planning. The Congressional Budget Offi

  ce, for instance,

  reduced its ten- year forecast for average US annual labor productivity

  growth from 1.8 percent in 2016 (CBO 2016) to 1.5 percent in 2017 (CBO

  2017). Although perhaps modest on its surface, that drop implies US gross

  domestic product (GDP) will be considerably smaller ten years from now

  than it would in the more optimistic scenario—a diff erence equivalent to

  almost $600 billion in 2017.

  1.3 Potential Explanations for the Paradox

  There are four principal candidate explanations for the current confl uence

  of technological optimism and poor productivity performance: (a) false

  hopes, (b) mismeasurement, (c) concentrated distribution and rent dissipa-

  tion, and (d) implementation and restructuring lags.7

  1.3.1 False

  Hopes

  The simplest possibility is that the optimism about the potential tech-

  nologies is misplaced and unfounded. Perhaps these technologies won’t be

  as transformative as many expect, and although they might have modest

  and noteworthy eff ects on specifi c sectors, their aggregate impact might be

  small. In this case, the paradox will be resolved in the future because realized

  productivity growth never escapes its current doldrums, which will force the

  optimists to mark their beliefs to market.

  History and some current examples off er a quantum of credence to this

  possibility. Certainly one can point to many prior exciting technologies that

  did not live up to initially optimistic expectations. Nuclear power never

  became too cheap to meter, and fusion energy has been twenty years away

  for sixty years. Mars may still beckon, but it has been more than forty years

  since Eugene Cernan was the last person to walk on the moon. Flying cars

  never got off the ground,8 and passenger jets no longer fl y at supersonic

  speeds. Even AI, perhaps the most promising technology of our era, is

  well behind Marvin Minsky’s 1967 prediction that “Within a generation

  the problem of creating ‘artifi cial intelligence’ will be substantially solved”

  (Minsky 1967, 2).

  On the other hand, there remains a compelling case for optimism. As we

  outline below, it is not diffi

  cult to construct back- of-the- envelope scenarios

  7. To some extent, these explanations parallel the explanations for the Solow paradox (Brynjolfsson 1993).

  8. But coming soon? https:// kittyhawk .aero / about/.

  Artifi cial Intelligence and the Modern Productivity Paradox 29

  in which even a modest number of currently existing technologies could

  combine to substantially raise productivity growth and societal welfare.

  Indeed, knowledgeable investors and researchers are betting their money

  and time on exactly such outcomes. Thus, while we recognize the potential

  for overoptimism—and the experience with early predictions for AI makes

  an especially relevant reminder for us to be somewhat circumspect in this

  chapter—we judge that it would be highly preliminary to dismiss optimism

  at this point.

  1.3.2 Mismeasurement

  Another potential explanation for the paradox is mismeasurement of out-

  put and productivity. In this case, it is the pessimistic reading of the empirical

  past, not the optimism about the future, that is mistaken. Indeed, this expla-

  nation implies that the productivity benefi ts of the new wave of technologies

  are already being enjoyed, but have yet to be accurately measured. Under

  this explanation, the slowdown of the past decade is illusory. This “mis-

  measurement hypothesis” has been put forth in several works (e.g., Mokyr

  2014; Alloway 2015; Feldstein 2015; Hatzius and Dawsey 2015; Smith 2015).

  There is a prima facie case for the mismeasurement hypothesis. Many new

  technologies, like smartphones, online social networks, and downloadable

  media involve little monetary cost, yet consumers spend large amounts of

  time with these technologies. Thus, the technologies might deliver substan-

  tial utility even if they account for a small share of GDP due to their low

  relative price. Guvenen et al. (2017) also show how growing off shore profi t

  shifting can be another source of mismeasurement.

  However, a set of recent studies provide good reason to think that mis-

  measurement is not the entire, or even a substantial, explanation for the

  slowdown. Cardarelli and Lusinyan (2015), Byrne, Fernald, and Reinsdorf

  (2016), Nakamura and Soloveichik (2015), and Syverson (2017), each using

  diff erent methodologies and data, present evidence that mismeasurement is

  not the primary explanation for the productivity slowdown. After all, while

  there is convincing evidence that many of the benefi ts of today’s technologies

  are not refl ected in GDP and therefore productivity statistics, the same was

  undoubtedly true in earlier eras as well.

  1.3.3 Concentrated Distribution and Rent Dissipation

  A third possibility is that the gains of the new technologies are already

  attainable, but that through a combination of concentrated distribution of

  those gains and dissipative eff orts to attain or preserve them (assuming the

  technologies are at least partially rivalrous), their eff ect on average produc-

  tivity growth is modest overall, and is virtually nil for the median worker. For

  instance, two of the most profi table uses of AI to date have been for targeting

  and pricing online ads, and for automated trading of fi nancial instruments,

  both applications with many zero- sum aspects.

  30 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  One version of this story asserts that the benefi ts of the new technologies

  are being enjoyed by a relatively small fraction of the economy, but the

  technologies’ narrowly scoped and rivalrous nature creates wasteful “gold

  rush”- type activities. Both those seeking to be one of the few benefi ciaries,

  as well as those who have attained some gains and seek to block access to

  others, engage in these dissipative eff orts, destroying many of the benefi ts

  of the new technologies.9

 
Recent research off ers some indirect support for elements of this story.

  Productivity diff erences between frontier fi rms and average fi rms in the same

  industry have been increasing in recent years (Andrews, Criscuolo, and Gal

  2016; Furman and Orszag 2015). Diff erences in profi t margins between the

  top and bottom performers in most industries have also grown (McAfee

  and Brynjolfsson 2008). A smaller number of superstar fi rms are gaining

  market share (Autor et al. 2017; Brynjolfsson et al. 2008), while workers’

  earnings are increasingly tied to fi rm- level productivity diff erences (Song

  et al. 2015). There are concerns that industry concentration is leading to sub-

  stantial aggregate welfare losses due to the distortions of market power (e.g.,

  De Loecker and Eeckhout 2017; Gutiérrez and Philippon 2017). Further-

  more, growing inequality can lead to stagnating median incomes and associ-

  ated socioeconomic costs, even when total income continues to grow.

  Although this evidence is important, it is not dispositive. The aggregate

  eff ects of industry concentration are still under debate, and the mere fact that

  a technology’s gains are not evenly distributed is no guarantee that resources

  will be dissipated in trying to capture them—especially that there would be

  enough waste to erase noticeable aggregate benefi ts.

  1.3.4 Implementation and Restructuring Lags

  Each of the fi rst three possibilities, especially the fi rst two, relies on ex-

  plaining away the discordance between high hopes and disappointing statis-

  tical realities. One of the two elements is presumed to be somehow “wrong.”

  In the misplaced optimism scenario, the expectations for technology by tech-

  nologists and investors are off base. In the mismeasurement explanation, the

  tools we use to gauge empirical reality are not up to the task of accurately

  doing so. And in the concentrated distribution stories, the private gains for

  the few may be very real, but they do not translate into broader gains for

  the many.

  But there is a fourth explanation that allows both halves of the seeming

  paradox to be correct. It asserts that there really is good reason to be optimis-

  tic about the future productivity growth potential of new technologies, while

  at the same time recognizing that recent productivity growth has been low.

  The core of this story is that it takes a considerable time—often more than

  9. Stiglitz (2014) off ers a diff erent mechanism where technological progress with concentrated benefi ts in the presence of restructuring costs can lead to increased inequality and even, in the short run, economic downturns.

  Artifi cial Intelligence and the Modern Productivity Paradox 31

  is commonly appreciated—to be able to suffi

  ciently harness new technolo-

  gies. Ironically, this is especially true for those major new technologies that

  ultimately have an important eff ect on aggregate statistics and welfare. That

  is, those with such broad potential application that they qualify as general

  purpose technologies (GPTs). Indeed, the more profound and far- reaching

  the potential restructuring, the longer the time lag between the initial inven-

  tion of the technology and its full impact on the economy and society.

  This explanation implies there will be a period in which the technologies

  are developed enough that investors, commentators, researchers, and policy-

  makers can imagine their potentially transformative eff ects, even though

  they have had no discernable eff ect on recent productivity growth. It isn’t

  until a suffi

  cient stock of the new technology is built and the necessary

  invention of complementary processes and assets occurs that the promise

  of the technology actually blossoms in aggregate economic data. Investors

  are forward looking and economic statistics are backward looking. In times

  of technological stability or steady change (constant velocity), the disjoint

  measurements will seem to track each other. But in periods of rapid change,

  the two measurements can become uncorrelated.

  There are two main sources of the delay between recognition of a new

  technology’s potential and its measurable eff ects. One is that it takes time

  to build the stock of the new technology to a size suffi

  cient enough to have

  an aggregate eff ect. The other is that complementary investments are neces-

  sary to obtain the full benefi t of the new technology, and it takes time to

  discover and develop these complements and to implement them. While the

  fundamental importance of the core invention and its potential for society

  might be clearly recognizable at the outset, the myriad necessary coinven-

  tions, obstacles, and adjustments needed along the way await discovery over

  time, and the required path may be lengthy and arduous. Never mistake a

  clear view for a short distance.

  This explanation resolves the paradox by acknowledging that its two

  seemingly contradictory parts are not actually in confl ict. Rather, both parts

  are in some sense natural manifestations of the same underlying phenom-

  enon of building and implementing a new technology.

  While each of the fi rst three explanations for the paradox might have a

  role in describing its source, the explanations also face serious questions

  in their ability to describe key parts of the data. We fi nd the fourth—the

  implementation and restructuring lags story—to be the most compelling in

  light of the evidence we discuss below. Thus it is the focus of our explorations

  in the remainder of this chapter.

  1.4 The Argument in Favor of the Implementation

  and Restructuring Lags Explanation

  Implicit or explicit in the pessimistic view of the future is that the recent slow-

  down in productivity growth portends slower productivity growth in the fu ture.

  32 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  We begin by establishing one of the most basic elements of the story: that

  slow productivity growth today does not rule out faster productivity growth

  in the future. In fact, the evidence is clear that it is barely predictive at all.

  Total factor productivity growth is the component of overall output

  growth that cannot be explained by accounting for changes in observable

  labor and capital inputs. It has been called a “measure of our ignorance”

  (Abramovitz 1956). It is a residual, so an econometrician should not be

  surprised if it is not very predictable from past levels. Labor productivity

  is a similar measure, but instead of accounting for capital accumulation,

  simply divides total output by the labor hours used to produce that output.

  Figures 1.3 and 1.4 plot, respectively, US productivity indices since 1948

  and productivity growth by decade. The data include average labor produc-

  tivity (LP), average total factor productivity (TFP), and Fernald’s (2014)

  utilization- adjusted TFP (TFPua).10

  Productivity has consistently grown in the postwar era, albeit at diff erent

  rates at diff erent times. Despite the consistent growth, however, past pro-

  ductivity growth rates have historically been poor predictors of future pro-

  ductivity growth. In other words, the productivity growth of the past
decade

  tells us little about productivity growth for the coming decade. Looking

  only at productivity data, it would have been hard to predict the decrease

  in productivity growth in the early 1970s or foresee the benefi cial impact of

  IT in the 1990s.

  As it turns out, while there is some correlation in productivity growth rates

  over short intervals, the correlation between adjacent ten- year periods is not

  statistically signifi cant. We present below the results from a regression of

  diff erent measures of average productivity growth on the previous period’s

  average productivity growth for ten- year intervals as well as scatterplots

  of productivity for each ten- year interval against the productivity in the

  subsequent period. The regressions in table 1.1 allow for autocorrelation

  in error terms across years (1 lag). Table 1.2 clusters the standard errors by

  decade. Similar results allowing for autocorrelation at longer time scales are

  presented in the appendix.

  In all cases, the R 2 of these regressions is low, and the previous decade’s

  productivity growth does not have statistically discernable predictive power

  over the next decade’s growth. For labor productivity, the R 2 is 0.009.

  Although the intercept in the regression is signifi cantly diff erent from zero

  (productivity growth is positive, on average), the coeffi

  cient on the previous

  period’s growth is not statistically signifi cant. The point estimate is economi-

  cally small, too. Taking the estimate at face value, 1 percent higher annual

  labor productivity growth in the prior decade (around an unconditional

  mean of about 2 percent per year) corresponds to less than 0.1 percent

  10. Available at http:// www .frbsf .org / economic - research / indicators - data / total - factor

  - productivity - tfp/.

  Fig. 1.3 US TFP and labor productivity indices, 1948– 2016

  Note: 1990 = 100.

  Fig. 1.4 US TFP and labor productivity growth (percent) by decade

  Table 1.1

  Regressions with Newey- West standard errors

  (1)

  (2)

  Labor

  Total factor

  (3)

  Newey- West regressions (1 lag allowed)

  productivity

  productivity

  Utilization- adjusted

  ten- year average productivity growth