The Economics of Artificial Intelligence Read online

  agenda. Perhaps more than anything, this volume highlights all that we do

  not know. It emphasizes questions around growth, inequality, privacy, trade,

  innovation, political economy, and so forth. We do not have answers yet. Of

  course, the lack of answers is a consequence of the early stage of AI’s diff u-

  sion. We cannot measure the impact until AI is widespread.

  With the current state of measurement, however, we may never get

  answers. As highlighted in the chapter by Raj and Seamans, we do not have

  18 Ajay Agrawal, Joshua Gans, and Avi Goldfarb

  good measures of AI. We also do not have a good measure of improvement

  to AI. What is the AI equivalent to the computational speed of a micro-

  chip or the horsepower of an internal combustion engine that will allow

  for quality- adjusted prices and measurement? We also do not have good

  measures of productivity growth when that growth is primarily driven by

  intangible capital. To answer these questions, the gross domestic product

  (GDP) measurement apparatus needs to focus on adjusting for intangible

  capital, software, and changes to the innovation process (Haskel and West-

  lake 2017). Furthermore, to the extent that the benefi ts of AI generate het-

  erogeneous benefi ts to people as consumers and as workers, measurement of

  the benefi t of AI will be tricky. For example, if AI enables more leisure and

  people choose to take more leisure, should that be accounted for in measures

  of inequality? If so, how?

  While each chapter has its own take on the agenda, several themes cut

  across the volume as key aspects of the research agenda going forward. To

  the extent there is consensus on the questions, the consensus focuses on the

  potential of AI as a GPT, and the associated potential consequences on

  growth and inequality. A second consistent theme is the role of regulation in

  accelerating or constraining the diff usion of the technology. A third theme is

  that AI will change the way we do our work as economists. Finally, a number

  of issues appear in many chapters that are somewhat outside the standard

  economic models of technology’s impact. How do people fi nd meaning if

  AI replaces work with leisure? How can economists inform the policy debate

  on solutions proposed by technologists in the popular press such as taxing

  robots or a universal basic income? How does a technology’s diff usion aff ect

  the political environment, and vice versa?

  This book highlights the questions and provides direction. We hope read-

  ers of this book take it as a starting point for their own research into this

  new and exciting area of study.

  References

  Agrawal, Ajay, Joshua Gans, and Avi Goldfarb. 2018. Prediction Machines: The

  Simple Economics of Artifi cial Intelligence. Boston, MA: Harvard Business Review Press.

  Akerman, Anders, Ingvil Gaarder, and Magne Mogstad. 2015. “The Skill Com-

  plementarity of Broadband Internet.” Quarterly Journal of Economics 130 (4):

  1781– 824.

  Arrow, Kenneth. (1951) 1963. Social Choice and Individual Values, 2nd ed. New York: John Wiley and Sons.

  Autor, David, David Dorn, Lawrence F. Katz, Christina Patterson, and John Van

  Reenen. 2017. “The Fall of the Labor Share and the Rise of Superstar Firms.”

  Working paper, Massachusetts Institute of Technology.

  Autor, David H., Lawrence F. Katz, and Alan B. Krueger. 1998. “Computing

  19

  Inequality: Have Computers Changed the Labor Market?” Quarterly Journal of

  Economics 113 (4): 1169– 213.

  Bostrom, Nick. 2014. Superintelligence: Paths, Dangers, Strategies. Oxford: Oxford University Press.

  Bresnahan, Timothy F., and M. Trajtenberg. 1995. “General Purpose Technologies

  ‘Engines of Growth’?” Journal of Econometrics 65:83– 108.

  David, Paul A. 1990. “The Dynamo and the Computer: An Historical Perspective

  on the Modern Productivity Paradox.” American Economic Review Papers and

  Proceedings 80 (2): 355– 61.

  Haskel, Jonathan, and Stian Westlake. 2017. Capitalism without Capital: The Rise of the Intangible Economy. Princeton, NJ: Princeton University Press.

  Kaplan, Jerry. 2016. Artifi cial Intelligence: What Everyone Needs to Know. Oxford: Oxford University Press.

  Kleinberg, Jon, Jens Ludwig, Sendhil Mullainathan, and Ziad Obermeyer. 2015.

  “Prediction Policy Problems.” American Economic Review 105 (5): 491– 95.

  I

  AI as a GPT

  1

  Artifi cial Intelligence and the

  Modern Productivity Paradox

  A Clash of Expectations

  and Statistics

  Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  The discussion around the recent patterns in aggregate productivity growth

  highlights a seeming contradiction. On the one hand, there are astonishing

  examples of potentially transformative new technologies that could greatly

  increase productivity and economic welfare (see Brynjolfsson and McAfee

  2014). There are some early concrete signs of these technologies’ promise,

  recent leaps in artifi cial intelligence (AI) performance being the most promi-

  nent example. However, at the same time, measured productivity growth

  over the past decade has slowed signifi cantly. This deceleration is large, cut-

  ting productivity growth by half or more in the decade preceding the slow-

  down. It is also widespread, having occurred throughout the Organisation

  for Economic Co- operation and Development (OECD) and, more recently,

  among many large emerging economies as well (Syverson 2017).1

  Erik Brynjolfsson is director of the MIT Initiative on the Digital Economy, the Schussel Family Professor of Management Science and professor of information technology at the MIT

  Sloan School of Management, and a research associate of the National Bureau of Economic Research. Daniel Rock is a PhD candidate at the MIT Sloan School of Management and a

  researcher at the MIT Initiative on the Digital Economy. Chad Syverson is the Eli B. and Har-riet B. Williams Professor of Economics at the University of Chicago Booth School of Business and a research associate of the National Bureau of Economic Research.

  We thank Eliot Abrams, Ajay Agrawal, David Autor, Seth Benzell, Joshua Gans, Avi Gold-

  farb, Austan Goolsbee, Andrea Meyer, Guillaume Saint- Jacques, Manuel Tratjenberg, and

  numerous participants at the NBER Workshop on AI and Economics in September 2017. In

  particular, Rebecca Henderson provided detailed and very helpful comments on an earlier draft and Larry Summers suggested the analogy to the J- curve. Generous funding for this research was provided in part by the MIT Initiative on the Digital Economy. This is a minor revision of NBER Working Paper no. 24001. For acknowledgments, sources of research support, and disclosure of the authors’ material fi nancial relationships, if any, please see http:// www .nber

  .org / chapters / c14007 .ack.

  1. A parallel, yet more pessimistically oriented debate about potential technological progress is the active discussion about robots taking jobs from more and more workers (e.g., Brynjolfsson and McAfee 2011; Acemoglu and Restrepo 2017; Bessen 2017; Autor and Salomons 2017).

  23

  24 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  We thus appear to be facing a redux of the Solow (1987) paradox: we

  see transformative new technologies everywhere but in the productivity sta-

  tistic
s.

  In this chapter, we review the evidence and explanations for the modern

  productivity paradox and propose a resolution. Namely, there is no inher-

  ent inconsistency between forward- looking technological optimism and

  backward- looking disappointment. Both can simultaneously exist. Indeed,

  there are good conceptual reasons to expect them to simultaneously exist

  when the economy undergoes the kind of restructuring associated with

  transformative technologies. In essence, the forecasters of future company

  wealth and the measurers of historical economic performance show the

  greatest disagreement during times of technological change. In this chap-

  ter, we argue and present some evidence that the economy is in such a

  period now.

  1.1 Sources of Technological Optimism

  Paul Polman, Unilever’s CEO, recently claimed that “The speed of inno-

  vation has never been faster.” Similarly, Bill Gates, Microsoft’s cofounder,

  observes that “Innovation is moving at a scarily fast pace.” Vinod Khosla of

  Khosla Ventures sees “the beginnings of . . . [a] rapid acceleration in the next

  10, 15, 20 years.” Eric Schmidt of Alphabet Inc., believes “we’re entering . . .

  the age of abundance [and] during the age of abundance, we’re going to see

  a new age . . . the age of intelligence.”2 Assertions like these are especially

  common among technology leaders and venture capitalists.

  In part, these assertions refl ect the continuing progress of information

  technology (IT) in many areas, from core technology advances like further

  doublings of basic computer power (but from ever larger bases) to suc-

  cessful investment in the essential complementary innovations like cloud

  infrastructure and new service- based business models. But the bigger source

  of optimism is the wave of recent improvements in AI, especially machine

  learning (ML). Machine learning represents a fundamental change from the

  fi rst wave of computerization. Historically, most computer programs were

  created by meticulously codifying human knowledge, mapping inputs to

  outputs as prescribed by the programmers. In contrast, machine- learning

  systems use categories of general algorithms (e.g., neural networks) to fi g-

  ure out relevant mappings on their own, typically by being fed very large

  sample data sets. By using these machine- learning methods that leverage

  the growth in total data and data- processing resources, machines have made

  impressive gains in perception and cognition, two essential skills for most

  2. http:// www .khoslaventures .com / fi reside - chat - with - google - co - founders - larry - page - and

  - sergey - brin; https:// en .wikipedia .org / wiki / Predictions _made _by _Ray _Kurzweil #2045: _The _Singularity; https:// www .theguardian .com / small - business - network / 2017 / jun / 22 / alphabets

  - eric - schmidt - google - artifi cial - intelligence - viva - technology - mckinsey.

  Artifi cial Intelligence and the Modern Productivity Paradox 25

  Fig. 1.1 AI versus human image recognition error rates

  types of human work. For instance, error rates in labeling the content of

  photos on ImageNet, a data set of over ten million images, have fallen from

  over 30 percent in 2010 to less than 5 percent in 2016, and most recently

  as low as 2.2 percent with SE- ResNet152 in the ILSVRC2017 competition

  (see fi gure 1.1).3 Error rates in voice recognition on the Switchboard speech

  recording corpus, often used to measure progress in speech recognition,

  have decreased to 5.5 percent from 8.5 percent over the past year (Saon et al.

  2017). The 5 percent threshold is important because that is roughly the per-

  formance of humans on each of these tasks on the same test data.

  Although not at the level of professional human performance yet, Face-

  book’s AI research team recently improved upon the best machine language

  translation algorithms available using convolutional neural net sequence

  prediction techniques (Gehring et al. 2017). Deep learning techniques have

  also been combined with reinforcement learning, a powerful set of tech-

  niques used to generate control and action systems whereby autonomous

  agents are trained to take actions given an environment state to maximize

  future rewards. Though nascent, advances in this fi eld are impressive. In

  addition to its victories in the game of Go, Google DeepMind has achieved

  superhuman performance in many Atari games (Fortunato et al. 2017).

  These are notable technological milestones. But they can also change the

  economic landscape, creating new opportunities for business value creation

  and cost reduction. For example, a system using deep neural networks was

  tested against twenty- one board- certifi ed dermatologists and matched their

  3. http:// image - net .org / challenges / LSVRC / 2017 / results. ImageNet includes labels for each image, originally provided by humans. For instance, there are 339,000 labeled as fl owers, 1,001,000 as food, 188,000 as fruit, 137,000 as fungus, and so on.

  26 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  performance in diagnosing skin cancer (Esteva et al. 2017). Facebook uses

  neural networks for over 4.5 billion translations each day.4

  An increasing number of companies have responded to these opportuni-

  ties. Google now describes its focus as “AI fi rst,” while Microsoft’s CEO

  Satya Nadella says AI is the “ultimate breakthrough” in technology. Their

  optimism about AI is not just cheap talk. They are making heavy invest-

  ments in AI, as are Apple, Facebook, and Amazon. As of September 2017,

  these companies comprise the fi ve most valuable companies in the world.

  Meanwhile, the tech- heavy NASDAQ composite index more than doubled

  between 2012 and 2017. According to CBInsights, global investment in

  private companies focused on AI has grown even faster, increasing from

  $589 million in 2012 to over $5 billion in 2016.5

  1.2 The Disappointing Recent Reality

  Although the technologies discussed above hold great potential, there is

  little sign that they have yet aff ected aggregate productivity statistics. Labor

  productivity growth rates in a broad swath of developed economies fell in

  the middle of the fi rst decade of the twenty- fi rst century and have stayed

  low since then. For example, aggregate labor productivity growth in the

  United States averaged only 1.3 percent per year from 2005 to 2016, less

  than half of the 2.8 percent annual growth rate sustained from 1995 to

  2004. Fully twenty- eight of the twenty- nine other countries for which the

  OECD has compiled productivity growth data saw similar decelerations.

  The unweighted average annual labor productivity growth rate across these

  countries was 2.3 percent from 1995 to 2004, but only 1.1 percent from 2005

  to 2015.6 What’s more, real median income has stagnated since the late 1990s

  and noneconomic measures of well- being, like life expectancy, have fallen

  for some groups (Case and Deaton 2017).

  Figure 1.2 replicates the Conference Board’s analysis of its country- level

  Total Economy Database (Conference Board 2016). It plots highly smoothed

  annual productivity growth rate series for the United States, other mature

  economies (which combined match mu
ch of the OECD sample cited above),

  emerging and developing economies, and the world overall. The aforemen-

  tioned slowdowns in the United States and other mature economies are clear

  in the fi gure. The fi gure also reveals that the productivity growth acceleration

  in emerging and developing economies during the fi rst decade of the twenty-

  4. https:// code .facebook .com / posts / 289921871474277 / transitioning - entirely - to-neural

  - machine- translation/.

  5. And the number of deals increased from 160 to 658. See https:// www .cbinsights .com

  / research / artifi cial - intelligence - startup - funding/.

  6. These slowdowns are statistically signifi cant. For the United States, where the slowdown is measured using quarterly data, equality of the two periods’ growth rates is rejected with a t- statistic of 2.9. The OECD numbers come from annual data across the thirty countries. Here, the null hypothesis of equality is rejected with a t- statistic of 7.2.

  Artifi cial Intelligence and the Modern Productivity Paradox 27

  Fig. 1.2 Smoothed average annual labor productivity growth (percent) by region

  Source: The Conference Board Total Economy DatabaseTM (adjusted version), November 2016.

  Note: Trend growth rates are obtained using HP fi lter, assuming a 1 = 100.

  fi rst century ended around the time of the Great Recession, causing a recent

  decline in productivity growth rates in these countries too.

  These slowdowns do not appear to simply refl ect the eff ects of the Great

  Recession. In the OECD data, twenty- eight of the thirty countries still

  exhibit productivity decelerations if 2008– 2009 growth rates are excluded

  from the totals. Cette, Fernald, and Mojon (2016), using other data, also fi nd

  substantial evidence that the slowdowns began before the Great Recession.

  Both capital deepening and total factor productivity (TFP) growth lead

  to labor productivity growth, and both seem to be playing a role in the slow-

  down (Fernald 2014; OECD 2015). Disappointing technological progress

  can be tied to each of these components. Total factor productivity directly

  refl ects such progress. Capital deepening is indirectly infl uenced by techno-

  logical change because fi rms’ investment decisions respond to improvements