The Economics of Artificial Intelligence
Page 95
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