Book Read Free

The Economics of Artificial Intelligence

Page 95

by Ajay Agrawal


  Data can also be created via a survey of fi rms. Survey data allows for

  more detailed inquiry than a census and can be carried out in a quicker

  and less expensive fashion. Further, a variety of organizations, both private

  and public, may have the interest and ability to conduct a survey regard-

  ing the adoption of AI or robotic technology. However, surveys introduce

  issues regarding sample selection and response rates, and depending on what

  organization is administering the survey, access to data can be limited or

  expensive.

  Collecting survey data regarding the adoption of technology is not an

  entirely new concept. The Survey of Manufacturing Technology (SMT) was

  conducted by the Census Bureau in collaboration with the Department of

  Defense in 1988, 1991, and 1993 to measure the diff usion, use, and planned

  future use of new technologies in the manufacturing sector of the United

  States. The SMT surveyed 10,000 establishments to learn about plant char-

  acteristics and adoption of seventeen established technologies grouped into

  fi ve categories: design and engineering, fabricated machining and assembly,

  automated material handling, automated sensors, and communication and

  control. Because the survey was administered by the Census Bureau, data

  from the SMT could easily be integrated with other fi rm- level data from

  the BLS or Census Bureau. The survey also allowed for panel analysis, as

  a subset of fi rms within the sample were respondents in multiple editions.

  Following the 1993 SMT, the Census Bureau discontinued the survey for

  funding reasons.

  562 Manav Raj and Robert Seamans

  The Department of Defense used the SMT data to assess the diff usion of

  technology, and other federal agencies used the data to gauge competitive-

  ness of the US manufacturing sector. The data were also used by the private

  sector in market analysis, competitiveness assessments, and planning. Mul-

  tiple academic studies, including Dunne (1994), Mcguckin, Streitwieser, and

  Doms (1998), Doms, Dunne, and Troske (1997) and Lewis (2011) analyzed

  the SMT data to address questions related to productivity growth, skill-

  biased technical change, earnings, and capital- labor substitution, among

  others.

  In many ways, the SMT could serve as a model for future inquiry into

  the adoption of robotics technology. It provided a broad look at the manu-

  facturing industry in the United States and allowed for the examination of

  eff ects over time and for fi rm- and individual- level analysis when integrated

  with other data from the BLS or Census Bureau. However, any updated ver-

  sion of the SMT would need to redefi ne the relevant technologies, examine

  the intensity of use, and investigate what tasks diff erent technologies are

  used for.

  Private data collected at individual fi rms can also be a useful tool. Inter-

  nal data from a fi rm exacerbates both the strengths and weaknesses of sur-

  vey data. Data collected at a single establishment can provide an unmatched

  level of detail and richness compared to data created by either a census or

  a survey. For example, Cowgill (2016) uses detailed individual- level skill

  and performance data from a single establishment to assess the returns

  to machine- learning algorithms used in hiring decisions. However, with a

  sample size of one, selection on fi rm is a highly salient issue and gener-

  alizability may be low. Further, any data produced will almost certainly

  be proprietary and diffi

  cult to get access to by other researchers, making

  reproducibility diffi

  cult (Lane 2003).

  22.6 Conclusion

  The recent dramatic increases in technological capabilities we have seen

  in the fi elds of robotics and artifi cial intelligence provide society with a

  myriad of opportunities and challenges. To eff ectively take advantage of

  these technologies, we must have a complete and thorough understanding

  of the impacts of these technologies on growth, productivity, labor, and

  equality. Systematic data on the adoption and use of these technologies,

  particularly at the establishment level, is necessary to understand the eff ects

  of these technologies on the economy and society as a whole. The creation

  and aggregation of these data sets through the census, surveys conducted

  by public or private organizations, and internal data collected at individual

  fi rms, would provide researchers and policymakers with the tools needed to

  empirically investigate the impact of these technologies, and craft appropri-

  ate responses to this phenomenon.

  AI, Labor, Productivity, and the Need for Firm- Level Data 563

  Finally, the need for high- quality data in this area is also linked with

  national competitiveness, particularly in relation to crafting appropriate

  policy responses. Mitchell and Brynjolfsson (2017) argue that the lack of

  information on AI could cripple our ability to prepare for the eff ects of

  technological advancement, leading to missed opportunities and poten-

  tially disastrous consequences. For example, decisions regarding whether

  to tax or subsidize AI or robots rely on understanding whether or not the

  particular technology serves as a substitute or complement to labor. These

  decisions can aff ect adoption patterns, and if made with an incomplete

  understanding of the eff ect of these technologies on labor markets, can lead

  to lower economic growth, less hiring, and lower wages. In addition, data

  must also be utilized to properly respond to consequences stemming from

  technology adoption. Identifying which populations may be most vulner-

  able to job displacement and eff ectively structuring job retraining programs

  requires a comprehensive understanding of the microlevel impacts of adop-

  tion of these technologies.

  References

  Acemoglu, Daron, and Pascual Restrepo. 2017. “Robots and Jobs: Evidence from

  US Labor Markets.” NBER Working Paper no. 23285, Cambridge, MA.

  Alexopoulos, Michelle, and Jon Cohen. 2016. “The Medium Is the Measure: Tech-

  nical Change and Employment, 1909– 1949.” Review of Economics and Statistics

  98 (4): 792– 810.

  Autor, David, and Anna Salomons. 2017. “Robocalypse Now— Does Productiv-

  ity Growth Threaten Employment?” Working paper, Massachusetts Institute of

  Technology.

  Bessen, James. 2017. “Automation and Jobs: When Technology Boosts Employ-

  ment.” Law and Economics Paper no. 17-09, Boston University School of Law.

  Bloom, Nicholas, Erik Brynjolfsson, Lucia Foster, Ron Jarmin, Megha Patnaik,

  Itay Saporta- Eksten, and John Van Reenen. 2017. “What Drives Diff erences in

  Management?” NBER Working Paper no. 23300, Cambridge, MA.

  Brynjolfsson, Erik, and Andrew McAfee. 2014. The Second Machine Age: Work,

  Progress, and Prosperity in a Time of Brilliant Technologies. New York: W. W.

  Norton.

  Brynjolfsson, Erik, and Kristina McElheran. 2016. “The Rapid Adoption of Data-

  Driven Decision- Making.” American Economic Review 106 (5): 133– 39.

  Brynjolfsson, Erik, and Tom Mitchell. 2017. “What Can Machine Learning Do?

  Workforce Implications.” Science 35
8 (6370): 1530– 34.

  Brynjolfsson, Erik, Tom Mitchell, and Daniel Rock. Forthcoming. “What Can

  Machines Learn, and What Does It Mean for Occupations and the Economy?”

  American Economic Association Papers and Proceedings.

  Brzeski, Carsten, and Inga Burk. 2015. “Die Roboter Kommen.” (“The Robots

  Come.”) ING DiBa Economic Research. https://

  www

  .ing-

  diba.de/

  binaries

  / content/ assets/ pdf/ ueber- uns/ presse/ publikationen/ ing- diba- economic- analysis _roboter- 2.0 .pdf.

  564 Manav Raj and Robert Seamans

  Council of Economic Advisers (CEA). 2016. “Economic Report of the President.”

  https:// obamawhitehouse.archives .gov/ administration/ eop/ cea/ economic- report

  - of-the- President/ 2016.

  Cowgill, Bo. 2016. “The Labor Market Eff ects of Hiring through Machine Learn-

  ing.” Working paper, Columbia University.

  Dauth, Wolfgang, Sebastian Findeisen, Jens Südekum, and Nicole Wößner. 2017.

  “German Robots—The Impact of Industrial Robots on Workers.” IAB Discus-

  sion Paper, Institut für Arbeitsmarkt- und Berufsforschung. https:// www .iab.de

  / en/ publikationen/ discussionpaper .aspx.

  Doms, Mark, Timothy Dunne, and Kenneth R. Troske. 1997. “Workers, Wages and

  Technology.” Quarterly Journal of Technology 62 (1): 253– 90.

  Dunne, Timothy. 1994. “Plant Age and Technology Use in U.S. Manufacturing

  Industries.” RAND Journal of Economics 25 (3): 488– 99.

  European Commission (EC). 2016. “Analysis of the Impact of Robotic Systems on

  Employment in the European Union—2012 Data Update.”

  Feldman, Maryann P., and Dieter F. Kogler. 2010. “Stylized Facts in the Geography

  of Innovation.” Handbook of the Economics of Innovation 1:381– 410.

  Felten, Ed, Manav Raj, and Rob Seamans. Forthcoming. “Linking Advances in

  Artifi cial Intelligence to Skills, Occupations, and Industries.” American Economics Association Papers & Proceedings.

  Frey, Carl B., and Michael A. Osborne. 2017. “The Future of Employment: How

  Susceptible Are Jobs to Computerisation?” Technological Forecasting and Social

  Change 114:254– 80.

  Furman, Jason. 2016a. “The Economic Case for Strengthening Unemployment

  Insurance.” Remarks at the Center for American Progress, Washington DC,

  July 11. https:// obamawhitehouse.archives .gov/ sites/ default/ files/ page/ files

  / 20160711_furman_uireform_cea .pdf.

  ———. 2016b. “Is This Time Diff erent? The Opportunities and Challenges of

  Artifi cial Intelligence.” Remarks at AI Now: The Social and Economic Implica-

  tions of Artifi cial Intelligence Technologies in the Near Term, New York Uni-

  versity, July 7. https:// obamawhitehouse.archives .gov/ sites/ default/ fi les/ page/ fi les

  / 20160707_cea_ai_furman .pdf.

  Graetz, Georg, and Guy Michaels. 2015. “Robots at Work.” CE P Discussion Paper

  no. 1335, Centre for Economic Performance.

  Lane, Julia. 2003. “Uses of Microdata: Keynote Speech.” In Statistical Confi den-

  tiality and Access to Microdata: Proceedings of the Seminar Session of the 2003

  Conference of European Statisticians, 11– 20. Geneva.

  Lewis, Ethan. 2011. “Immigration, Skill Mix, and Capital Skill Complementarity.”

  Quarterly Journal of Economics 126 (2): 1029– 69.

  Mandel, Michael. 2017. “How Ecommerce Creates Jobs and Reduces Income

  Inequality.” Working paper, Progressive Policy Institute. http:// www .progressive

  policy .org/ wp- content/ uploads/ 2017/ 09/ PPI_ECommerceInequality- fi nal .pdf.

  Mann, Katja, and Lukas Püttmann. 2017. “Benign Eff ects of Automation: New

  Evidence from Patent Texts.” Unpublished manuscript.

  Mcguckin, Robert H., Mary L. Streitwieser, and Mark Doms. 1998. “The Eff ect

  of Technology Use on Productivity Growth.” Economics of Innovation and New

  Technology 7 (1): 1– 26.

  McKinsey Global Institute (MGI). 2017. “Artifi cial Intelligence the Next Digital

  Frontier?” https:// www .mckinsey .com/ business- functions/ mckinsey- analytics

  / our- insights/ how- artifi cial- intelligence- can- deliver- real- value- to-companies.

  Mitchell, Tom, and Erik Brynjolfsson. 2017. “Track How Technology Is Transform-

  ing Work.” Nature 544 (7650): 290– 92.

  AI, Labor, Productivity, and the Need for Firm- Level Data 565

  National Academy of Sciences (NAS). 2017. “Information Technology and the U.S.

  Workforce: Where Are We and Where Do We Go from Here?” https:// www .nap

  .edu/ catalog/ 24649/ information- technology- and- the- us- workforce- where- are

  - we- and.

  Pajarinen, Mike, and Petri Rouvinen. 2014. “Computerization Threatens One Third

  of Finnish Employment.” ETLA Brief no. 22, Research Institute of the Finnish

  Economy.

  23

  How Artifi cial Intelligence and

  Machine Learning Can Impact

  Market Design

  Paul R. Milgrom and Steven Tadelis

  23.1 Introduction

  For millennia, markets have played a key role in providing individuals and

  businesses with the opportunity to gain from trade. More often than not,

  markets require structure and a variety of intuitional support to operate

  effi

  ciently. For example, auctions have become a commonly used mechanism

  to generate gains from trade when price discovery is essential. Research in

  the area now commonly referred to as market design, going back to Vickrey

  (1961), demonstrated that it is critical to design auctions and market institu-

  tions more broadly in order to achieve effi

  cient outcomes (see, e.g., Milgrom

  2017; Roth 2015).

  Any market designer needs to understand some fundamental details of

  the transactions that are expected to be consummated in order to design the

  most eff ective and effi

  cient market structure to support these transactions.

  For example, the National Resident Matching Program, which matches doc-

  tors to hospital residencies, was originally designed in an era when nearly all

  doctors were men and wives followed them to their residencies. It needed to

  be redesigned in the 1990s to accommodate the needs of couples, when men

  and women doctors could no longer be assigned jobs in diff erent cities. Even

  Paul R. Milgrom is the Shirley and Leonard Ely Professor of Humanities and Sciences in the Department of Economics at Stanford University and professor, by courtesy, at both the Department of Management Science and Engineering and the Graduate School of Business.

  Steven Tadelis holds the James J. and Marianne B. Lowrey Chair in Business and is professor of economics, business, and public policy at the University of California, Berkeley, and a research associate of the National Bureau of Economic Research.

  For acknowledgments, sources of research support, and disclosure of the authors’ material fi nancial relationships, if any, please see http:// www .nber .org/ chapters/ c14008.ack.

  567

  568 Paul R. Milgrom and Steven Tadelis

  something as mundane as the sale of a farm when a farmer dies requires

  knowledge of the structure and decisions about whether to sell the whole

  farm as a unit, or to separate the house for sale as a weekend retreat while

  selling the land to neighboring farmers, or selling the forest separately t
o a

  wildlife preservation fund.

  In complex environments, it can be diffi

  cult to understand the underlying

  characteristics of transactions, and it is challenging to learn enough about

  them in order to design the best institutions to effi

  ciently generate gains

  from trade. For example, consider the recent growth of online advertising

  exchanges that match advertisers with online ads. Many ads are allocated

  to advertisers using real- time auctions. But how should publishers design

  these auctions in order to make the best use of their advertising space, and

  how can they maximize the returns to their activities? Based on the early

  theoretical auction design work of Myerson (1981), Ostrovsky and Schwartz

  (2017) have shown that a little bit of market design in the form of setting

  better reserve prices can have a dramatic impact on the profi ts an online ad

  platform can earn.

  But how can market designers learn the characteristics necessary to set

  optimal, or at least better, reserve prices? Or, more generally, how can market

  designers better learn the environment of their markets? In response to these

  challenges, artifi cial intelligence (AI) and machine learning are emerging as

  important tools for market design. Retailers and marketplaces such as eBay,

  TaoBao, Amazon, Uber, and many others are mining their vast amounts of

  data to identify patterns that help them create better experiences for their

  customers and increase the effi

  ciency of their markets. By having better

  prediction tools, these and other companies can predict and better manage

  sophisticated and dynamic market environments. The improved forecasting

  that AI and machine- learning algorithms provide help marketplaces and

  retailers better anticipate consumer demand and producer supply as well as

  help target products and activities to fi ner segmented markets.

  Turning back to markets for online advertising, two- sided markets such

  as Google, which match advertisers with consumers, are not only using AI

  to set reserve prices and segment consumers into fi ner categories for ad tar-

  geting, but they also develop AI- based tools to help advertisers bid on ads.

  In April 2017 Google introduced “Smart Bidding,” a product based on AI

  and machine learning that helps advertisers bid automatically on ads based

 

‹ Prev