The Economics of Artificial Intelligence Read online

  tern. Only one self- driving vehicle needs to experience an anomaly for many

  vehicles to learn from it. Waymo, a subsidiary of Google, has cars driv-

  ing 25,000 “real” autonomous and about 19 million simulated miles each

  week.16 All of the Waymo cars learn from the joint experience of the others.

  Similarly, a robot struggling with a task can benefi t from sharing data and

  learnings with other robots that use a compatible knowledge- representation

  framework.17

  When one thinks of AI as a GPT, the implications for output and wel-

  fare gains are much larger than in our earlier analysis. For example, self-

  driving cars could substantially transform many nontransport industries.

  15. For example, through enterprise resource planning systems in factories, internet commerce, mobile phones, and the “Internet of Things.”

  16. http:// ben - evans .com / benedictevans / 2017 / 8 / 20 / winner - takes - all.

  17. Rethink Robotics is developing exactly such a platform.

  Artifi cial Intelligence and the Modern Productivity Paradox 41

  Retail could shift much further toward home delivery on demand, creating

  consumer welfare gains and further freeing up valuable high- density land

  now used for parking. Traffi

  c and safety could be optimized, and insurance

  risks could fall. With over 30,000 deaths due to automobile crashes in the

  United States each year, and nearly a million worldwide, there is an oppor-

  tunity to save many lives.18

  1.7 Why Future Technological Progress Is Consistent

  with Low Current Productivity Growth

  Having made a case for technological optimism, we now turn to explain-

  ing why it is not inconsistent with—and in fact may even be naturally related

  to—low current productivity growth.

  Like other GPTs, AI has the potential to be an important driver of

  productivity. However, as Jovanovic and Rousseau (2005) point out (with

  additional reference to David’s [1991] historical example), “a GPT does

  not deliver productivity gains immediately upon arrival” (1184). The tech-

  nology can be present and developed enough to allow some notion of its

  transformative eff ects even though it is not aff ecting current productivity

  levels in any noticeable way. This is precisely the state that we argue the

  economy may be in now.

  We discussed earlier that a GPT can at one moment both be present and

  yet not aff ect current productivity growth if there is a need to build a suf-

  fi ciently large stock of the new capital, or if complementary types of capital,

  both tangible and intangible, need to be identifi ed, produced, and put in

  place to fully harness the GPT’s productivity benefi ts.

  The time necessary to build a suffi

  cient capital stock can be extensive.

  For example, it was not until the late 1980s, more than twenty- fi ve years

  after the invention of the integrated circuit, that the computer capital stock

  reached its long- run plateau at about 5 percent (at historical cost) of total

  nonresidential equipment capital. It was at only half that level ten years

  prior. Thus, when Solow pointed out his now eponymous paradox, the com-

  puters were fi nally just then getting to the point where they really could be

  seen everywhere.

  David (1991) notes a similar phenomenon in the diff usion of electrifi ca-

  tion. At least half of US manufacturing establishments remained unelectri-

  fi ed until 1919, about thirty years after the shift to polyphase alternating

  current began. Initially, adoption was driven by simple cost savings in pro-

  18. These latter two consequences of autonomous vehicles, while certainly refl ecting welfare improvements, would need to be capitalized in prices of goods or services to be measured in standard GDP and productivity measures. We will discuss AI- related measurement issues in greater depth later. Of course, it is worth remembering that autonomous vehicles also hold the potential to create new economic costs if, say, the congestion from lower marginal costs of operating a vehicle is not counteracted by suffi

  ciently large improvements in traffi

  c management

  technology or certain infrastructure investments.

  42 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  viding motive power. The biggest benefi ts came later, when complementary

  innovations were made. Managers began to fundamentally reorganize work

  by replacing factories’ centralized power source and giving every individual

  machine its own electric motor. This enabled much more fl exibility in the

  location of equipment and made possible eff ective assembly lines of mate-

  rials fl ow.

  This approach to organizing factories is obvious in retrospect, yet it took

  as many as thirty years for it to become widely adopted. Why? As noted

  by Henderson (1993, 2006), it is exactly because incumbents are designed

  around the current ways of doing things and so profi cient at them that they

  are blind to or unable to absorb the new approaches and get trapped in the

  status quo—they suff er the “curse of knowledge.”19

  The factory electrifi cation example demonstrates the other contributor to

  the time gap between a technology’s emergence and its measured productiv-

  ity eff ects: the need for installation (and often invention) of complementary

  capital. This includes both tangible and intangible investments. The time-

  line necessary to invent, acquire, and install these complements is typically

  more extensive than the time- to-build considerations just discussed. Con-

  sider the measured lag between large investments in IT and productivity

  benefi ts within fi rms. Brynjolfsson and Hitt (2003) found that while small

  productivity benefi ts were associated with fi rms’ IT investments when one-

  year diff erences were considered, the benefi ts grew substantially as longer

  diff erences were examined, peaking after about seven years. They attributed

  this pattern to the need for complementary changes in business processes.

  For instance, when implementing large enterprise- planning systems, fi rms

  almost always spend several times more money on business process rede-

  sign and training than on the direct costs of hardware and software. Hiring

  and other human- resources practices often need considerable adjustment

  to match the fi rm’s human capital to the new structure of production. In

  fact, Bresnahan, Brynjolfsson, and Hitt (2002) fi nd evidence of three- way

  complementarities between IT, human capital, and organizational changes

  in the investment decisions and productivity levels. Furthermore, Brynjolfs-

  son, Hitt, and Yang (2002) show each dollar of IT capital stock is cor-

  related with about $10 of market value. They interpret this as evidence of

  substantial IT- related intangible assets and show that fi rms that combine IT

  investments with a specifi c set of organizational practices are not just more

  productive, they also have disproportionately higher market values than

  fi rms that invest in only one or the other. This pattern in the data is consistent

  with a long stream of research on the importance of organizational and even

  19. Atkeson and Kehoe (2007) note manufacturers’ reluctance to abandon their large knowledge stock at the beginning of the transition to electric
power to adopt what was, initially, only a marginally superior technology. David and Wright (2006) are more specifi c, focusing on “the need for organizational and above all for conceptual changes in the ways tasks and products are defi ned and structured” (147, emphasis in original).

  Artifi cial Intelligence and the Modern Productivity Paradox 43

  cultural change when making IT investments and technology investments

  more generally (e.g., Aral, Brynjolfsson, and Wu 2012; Brynjolfsson and

  Hitt 2000; Orlikowski 1996; Henderson 2006).

  But such changes take substantial time and resources, contributing to

  organizational inertia. Firms are complex systems that require an extensive

  web of complementary assets to allow the GPT to fully transform the sys-

  tem. Firms that are attempting transformation often must reevaluate and

  reconfi gure not only their internal processes but often their supply and distri-

  bution chains as well. These changes can take time, but managers and entre-

  preneurs will direct invention in ways that economize on the most expensive

  inputs (Acemoglu and Restrepo 2017). According to LeChatelier’s principle

  (Milgrom and Roberts 1996), elasticities will therefore tend to be greater in

  the long run than in the short run as quasi- fi xed factors adjust.

  There is no assurance that the adjustments will be successful. Indeed,

  there is evidence that the modal transformation of GPT- level magnitude

  fails. Alon et al. (2017) fi nd that cohorts of fi rms over fi ve years old con-

  tribute little to aggregate productivity growth on net—that is, among estab-

  lished fi rms, productivity improvements in one fi rm are off set by produc-

  tivity declines in other fi rms. It is hard to teach the proverbial old dog new

  tricks. Moreover, the old dogs (companies) often have internal incentives to

  not learn them (Arrow 1962; Holmes, Levine, and Schmitz 2012). In some

  ways, technology advances in industry one company death at a time.

  Transforming industries and sectors requires still more adjustment and

  reconfi guration. Retail off ers a vivid example. Despite being one of the

  biggest innovations to come out of the 1990s dot- com boom, the largest

  change in retail in the two decades that followed was not e-commerce, but

  instead the expansion of warehouse stores and supercenters (Hortaçsu

  and Syverson 2015). Only very recently did e-commerce become a force for

  general retailers to reckon with. Why did it take so long? Brynjolfsson and

  Smith (2000) document the diffi

  culties incumbent retailers had in adapting

  their business processes to take full advantage of the internet and electronic

  commerce. Many complementary investments were required. The sector

  as a whole required the build out of an entire distribution infrastructure.

  Customers had to be “retrained.” None of this could happen quickly. The

  potential of e-commerce to revolutionize retailing was widely recognized,

  and even hyped in the late 1990s, but its actual share of retail commerce was

  miniscule, 0.2 percent of all retail sales in 1999. Only after two decades of

  widely predicted yet time- consuming change in the industry, is e-commerce

  starting to approach 10 percent of total retail sales and companies like Ama-

  zon are having a fi rst- order eff ect on more traditional retailers’ sales and

  stock market valuations.

  The case of self- driving cars discussed earlier provides a more prospective

  example of how productivity might lag technology. Consider what happens

  to the current pools of vehicle production and vehicle operation workers

  44 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  when autonomous vehicles are introduced. Employment on production side

  will initially increase to handle research and development (R&D), AI de-

  velopment, and new vehicle engineering. Furthermore, learning curve issues

  could well imply lower productivity in manufacturing these vehicles during

  the early years (Levitt, List, and Syverson 2013). Thus labor input in the

  short run can actually increase, rather than decrease, for the same amount

  of vehicle production. In the early years of autonomous vehicle develop-

  ment and production, the marginal labor added by producers exceeds the

  marginal labor displaced among the motor vehicle operators. It is only later

  when the fl eet of deployed autonomous vehicles gets closer to a steady state

  that measured productivity refl ects the full benefi ts of the technology.

  1.8 Viewing Today’s Paradox through the Lens

  of Previous General Purpose Technologies

  We have indicated in the previous discussion that we see parallels between

  the current paradox and those that have happened in the past. It is closely

  related to the Solow paradox era circa 1990, certainly, but it is also tied

  closely to the experience during the diff usion of portable power (combining

  the contemporaneous growth and transformative eff ects of electrifi cation

  and the internal combustion engine).

  Comparing the productivity growth patterns of the two eras is instructive.

  Figure 1.8 is an updated version of an analysis from Syverson (2013). It over-

  lays US labor productivity since 1970 with that from 1890 to 1940, the period

  after portable power technologies had been invented and were starting to

  be placed into production. (The historical series values are from Kendrick

  [1961].) The modern series timeline is indexed to a value of 100 in 1995 and

  Fig. 1.8 Labor productivity growth in the portable power and IT eras

  Artifi cial Intelligence and the Modern Productivity Paradox 45

  is labeled on the upper horizontal axis. The portable power era index has a

  value of 100 in 1915, and its years are shown on the lower horizontal axis.

  Labor productivity during the portable power era shared remarkably

  similar patterns with the current series. In both eras, there was an initial

  period of roughly a quarter century of relatively slow productivity growth.

  Then both eras saw decade- long accelerations in productivity growth, span-

  ning 1915 to 1924 in the portable power era and 1995 to 2004 more recently.

  The late- 1990s acceleration was the (at least partial) resolution of the

  Solow paradox. We imagine that the late 1910s acceleration could have simi-

  larly answered some economist’s query in 1910 as to why one sees electric

  motors and internal combustion engines everywhere but in the productivity

  statistics.20

  Very interesting, and quite relevant to the current situation, the produc-

  tivity growth slowdown we have experienced after 2004 also has a parallel

  in the historical data, a slowdown from 1924 to 1932. As can be seen in the

  fi gure, and instructive to the point of whether a new wave of AI and associ-

  ated technologies (or if one prefers, a second wave of IT- based technology)

  could reaccelerate productivity growth, labor productivity growth at the end

  of the portable power era rose again, averaging 2.7 percent per year between

  1933 and 1940.

  Of course this past breakout growth is no guarantee that productivity

  must speed up again today. However, it does raise two relevant points. First,

  it is another example of a period of sluggish productivity growth followed

  by
an acceleration. Second, it demonstrates that productivity growth driven

  by a core GPT can arrive in multiple waves.

  1.9 Expected Productivity Eff ects of an AI- Driven Acceleration

  To understand the likely productivity eff ects of AI, it is useful to think

  of AI as a type of capital, specifi cally a type of intangible capital. It can be

  accumulated through investment, it is a durable factor of production, and

  its value can depreciate. Treating AI as a type of capital clarifi es how its

  development and installation as a productive factor will aff ect productivity.

  As with any capital deepening, increasing AI will raise labor productivity.

  This would be true regardless of how well AI capital is measured (which we

  might expect it won’t be for several reasons discussed below) though there

  may be lags.

  The eff ects of AI on TFP are more complex and the impact will depend

  on its measurement. If AI (and its output elasticity) were to be measured

  perfectly and included in both the input bundle in the denominator of TFP

  20. We are not aware of anyone who actually said this, and of course today’s system of national economic statistics did not exist at that time, but we fi nd the scenario amusing, instructive, and in some ways plausible.

  46 Erik Brynjolfsson, Daniel Rock, and Chad Syverson

  and the output bundle in the numerator, then measured TFP will accurately

  refl ect true TFP. In this case, AI could be treated just like any other measur-

  able capital input. Its eff ect on output could be properly accounted for and

  “removed” by the TFP input measure, leading to no change in TFP. This

  isn’t to say that there would not be productive benefi ts from diff usion of AI;

  it is just that it could be valued like other types of capital input.

  There are reasons why economists and national statistical agencies might

  face measurement problems when dealing with AI. Some are instances of

  more general capital measurement issues, but others are likely to be idiosyn-

  cratic to AI. We discuss this next.

  1.10 Measuring

  AI

  Capital

  Regardless of the eff ects of AI and AI- related technologies on actual out-

  put and productivity, it is clear from the productivity outlook that the ways

  AI’s eff ects will be measured are dependent on how well countries’ statistics