by Ajay Agrawal
Clark, Colin. 1940. The Conditions of Economic Progress. London: Macmillan.
Comin, Diego A., Danial Lashkari, and Martí Mestieri. 2015. “Structural Change
with Long- Run Income and Price Eff ects.” NBER Working Paper no. 21595,
Cambridge, MA.
Dennis, Benjamin N., and Talan B. I˙sçan. 2009. “Engel versus Baumol: Accounting
for Structural Change Using Two Centuries of US Data.” Explorations in Eco-
nomic History 46 (2): 186– 202.
Dupuit, Jules. 1844. “De la Mesure de L’utilité des Travaux Publics.” Annales des
Ponts et Chaussées 8 (2 sem): 332– 75.
Engel, Ernst. 1857. “Die Productions- und Consumtionsverhältnisse des König-
reichs Sachsen.” Zeitschrift des Statistischen Bureaus des Königlich Sächsischen
Ministerium des Inneren. 8– 9:28– 29.
Foellmi, Reto, and Josef Zweimüller. 2008. “Structural Change, Engel’s Consump-
tion Cycles and Kaldor’s Facts of Economic Growth.” Journal of Monetary Eco-
nomics 55 (7): 1317– 28.
Ford, Martin. 2015. Rise of the Robots: Technology and the Threat of a Jobless Future.
New York: Basic Books.
Frey, Carl Benedikt, and Michael A. Osborne. 2017. “The Future of Employment:
How Susceptible are Jobs to Computerisation?” Technological Forecasting and
Social Change 114 (2017): 254– 80.
Gaggl, Paul, and Greg C. Wright. 2017. “A Short- Run View of What Computers
Do: Evidence from a UK Tax Incentive.” American Economic Journal: Applied
Economics 9 (3): 262– 94.
Gordon, Robert D. 1941. “Values of Mills’ Ratio of Area to Bounding Ordinate and
of the Normal Probability Integral for Large Values of the Argument.” Annals of
Mathematical Statistics 12 (3): 364– 66.
Kollmeyer, Christopher. 2009. “Explaining Deindustrialization: How Affl
uence,
Productivity Growth, and Globalization Diminish Manufacturing Employment
1.” American Journal of Sociology 114 (6): 1644– 74.
Kongsamut, Piyabha, Sergio Rebelo, and Danyang Xie. 2001. “Beyond Balanced
Growth.” Review of Economic Studies 68 (4): 869– 82.
Artifi cial Intelligence and Jobs: The Role of Demand 307
Lawrence, Robert Z., and Lawrence Edwards. 2013. “US Employment Deindustri-
alization: Insights from History and the International Experience.” Policy Brief
no. 13– 27, Peterson Institute for International Economics.
Lewis, W. Arthur. 1954. “Economic Development with Unlimited Supplies of
Labour.” Manchester School 22 (2): 139– 91.
Matsuyama, Kiminori. 1992. “Agricultural Productivity, Comparative Advantage,
and Economic Growth.” Journal of Economic Theory 58 (2): 317– 34.
———. 2009. “Structural Change in an Interdependent World: A Global View of
Manufacturing Decline.” Journal of the European Economic Association 7 (2– 3):
478– 86.
———. 2002. “The Rise of Mass Consumption Societies.” Journal of Political
Economy 110 (5): 1035– 70.
Ngai, L. Rachel, and Christopher A. Pissarides. 2007. “Structural Change in a Multi-
sector Model of Growth.” American Economic Review 97 (1): 429– 43.
Nickell, Stephen, Stephen Redding, and Joanna Swaffi
eld. 2008. “The Uneven Pace
of Deindustrialisation in the OECD.” World Economy 31 (9): 1154– 84.
Parker, William N., and Judith L. V. Klein. 1966. “Productivity Growth in Grain
Production in the United States, 1840– 60 and 1900– 10.” In Output, Employment,
and Productivity in the United States after 1800, edited by Dorothy Brady, 523– 82.
Cambridge, MA: National Bureau of Economic Research.
Rodrik, Dani. 2016. “Premature Deindustrialization.” Journal of Economic Growth
21 (1): 1– 33.
Rowthorn, Robert, and Ramana Ramaswamy. 1999. “Growth, Trade, and Deindus-
trialization.” IMF Staff Papers 46 (1): 18– 41.
Tirole, Jean. 1988. The Theory of Industrial Organization. Cambridge, MA: MIT
Press.
United States Bureau of the Census. 1975. Historical Statistics of the United States, Colonial Times to 1970, no. 93. US Department of Commerce, Bureau of the
Census.
11
Public Policy in an AI Economy
Austan Goolsbee
11.1 Introduction
This conference has brought together a mix of technology and economics
scholars to think broadly about the role of artifi cial intelligence (AI) in the
economy, and this short chapter will present a few thoughts about the role
of policy in a world where AI becomes ubiquitous.
Most of the public discussion about an AI- dominated economy has
focused on robots and the future of work. Ruminations by public fi gures
like Bill Gates, Stephen Hawking, and Elon Musk have stoked fears that
robots will destroy our jobs (and, possibly, the world). Some of these same
fi gures have called for various heterodox policy ideas, too, from moving to
colonies in space to taxing the robots to providing a universal basic income
(UBI) untethered to work.
As the research and comments in this volume suggest, economists have
generally been less pessimistic when thinking about the role of AI on jobs.
They often highlight the historical record of job creation despite job dis-
placement, documented the way technological advances have eliminated
jobs in some sectors but expanded jobs and increased wages in the economy
overall, and highlighted the advantages that the new technologies will likely
have in the future (some recent discussions include Autor 2015; Autor and
Salomons 2018; Brynjolfsson and McAfee 2014; Mokyr 2014).
Austan Goolsbee is the Robert P. Gwinn Professor of Economics at the University of Chicago Booth School of Business and a research associate of the National Bureau of Economic Research.
I wish to thank the participants at the NBER Artifi cial Intelligence conference for helpful comments. For acknowledgments, sources of research support, and disclosure of the author’s or authors’ material fi nancial relationships, if any, please see http:// www .nber .org/ chapters
/ c14030.ack.
309
310 Austan Goolsbee
The pessimistic case has come more from technology/ business sector.
Perhaps seeing the advances in technology up close, they worry that the
machines may soon be so good that they could replace almost anyone. One
major study across many industries by the McKinsey Global Institute (2017)
argues that 73 million jobs may be destroyed by automation by 2030 because
of the rise of the new technologies.
In many ways, it is unfortunate that labor market policy has dominated
our thinking about the AI economy. The main economic impact of AI is
not about jobs or, at least, is about much more than just jobs. The main
economic impact of these technologies will be how good they are. If the
recent advances continue, AI has the potential to improve the quality of
our products and our standard of living. If AI helps us diagnose medical
problems better, improves our highway safety, gives us back hours of our
day that were spent driving in traffi
c, or even just improves the quality of
our selfi es, these are direct consumer benefi ts. These raise our real incomes
and the economic studies valuing the improv
ements from quality and from
new products tend to show their value is often extremely high (see the dis-
cussions in the volume of Bresnahan and Gordon [1997] or the discussions
over valuing “free” goods like Goolsbee and Klenow [2006] and Varian
[2013]).
That is a diff erent way of saying that if AI succeeds, it will raise our pro-
ductivity and higher productivity makes us rich. It is not a negative. Indeed,
if AI succeeded in the way some fear, it would mean the exact reversal of
the main problem facing growth in the last decade or more that productiv-
ity growth has been too slow. Indeed, it would decisively refute one of the
central tenets of secular stagnationist thinkers like Gordon (2016), who
argue that low productivity growth is a semi- permanent condition for the
advanced economies because of the scarcity of path breaking ideas. Would
that AI could change that equation.
This chapter will consider a few disparate thoughts about policy in an
AI- intensive economy (interpreting AI broadly to include a cluster of infor-
mation technology- based productivity improvements beyond just conven-
tional artifi cial intelligence or machine learning). It will consider the speed
of adoption of the technology—the impact on the job market and the impli-
cations for inequality across people and across places, discuss the challenges
of enacting a universal basic income as a response to widespread AI adop-
tion, discuss pricing, privacy and competition policy, and conclude with the
question of whether AI will improve policy making itself.
11.2 The Speed of Adoption: Implications
for the Job Market and for Inequality
Taking the issue of job displacement fi rst, the basic conclusion of the
economists is that for the last hundred years there have been massive
Public Policy in an AI Economy 311
amounts of job displacement, yet the structural unemployment rate has
not seemed to rise, much less trend toward 100 percent. Over time, people
adjust. They move. They get skills. The long- run impact of labor- saving
technologies has overwhelming been positive for market economies. If the
fear is that AI will replace low- skill jobs, it is a fact that tens or even hun-
dreds of millions of low- skill jobs were displaced by technology in previous
years in a process very similar to the one we describe today. If the fear is
that AI is diff erent this time around because it will begin to replace types of
jobs that have never been automated before like higher- skill or white- collar
jobs, the historical data indicate that those groups have been able to adjust
to shocks and move to new sectors and new geographic areas easier than
lower- skill workers have.
A critical issue is, of course, how fast the adjustment takes place/ the speed
of adoption of AI technology. The economy has proven quite capable of
inventing new things for people to do over the long run. Obviously, if change
happens all at once, the adjustment problem is worst. Spread out over time,
however, the adjustment can be manageable. Take the much discussed case
of autonomous cars. There were about 3.5 million truck, bus, and taxi driv-
ers in 2015, and suppose that every one of them were lost due to advances
in self- driving car technology. If this loss takes place over fi fteen years, this
would average a little over 19,000 per month, and compare that to the fact
that in 2017 the Job Openings and Labor Turnover Survey (JOLTS) data
show that the economy generated about 5.3 million jobs per month (with
5.1 million separations per month). The complete elimination of every job
in the sector would increase the separation rate by less than four- tenths of
a percent. It would force drivers into new sectors and be disruptive to their
livelihoods. But as a macroeconomic phenomenon, the impact would be
small. If that loss happened in two years, the impact would be quite signifi -
cant. So it is worth considering what infl uences the speed of adoption and,
certainly, a key determinant will be how good the AI actually is compared
to people. But, many analysts seem to view that as the only thing that will
determine adoption rates. It is worth considering at least two other factors:
prices and adjustment costs.
First, many of these AI innovations involve signifi cant capital outlays
up front and that alone may slow their adoption for some time. Ride- share
drivers, for example, by some measures can barely cover the cost of operat-
ing their cars (including depreciation, fuel, maintenance, and insurance) at
the price of cars now. AI- enabled autonomous vehicles are likely to cost
substantially more per car than conventional cars when they become avail-
able to the public. Will companies be willing to incur large upfront costs to
bypass paying drivers? It really depends on prices that we do not yet know.
Second, “better” does not always mean faster adoption. Economists
have shown automated stock picking through index funds superior to
active management for decades, yet people still hold trillions in ineffi
cient,
312 Austan Goolsbee
high- fee funds. Millions of people have mortgages with higher than market
interest rates that they do not refi nance, cell phone data plans that do not
match their usage, and so on. There are tens of millions of people that do
not use the internet. Inertia is a powerful force slowing the adoption of tech-
nology products and is certainly worth remembering if we want to predict
something like how fast people will give up common behaviors like driving
for themselves.
Third, in an important sense, we know that AI can only be as good as its
training sample and there are some very diff erent types of customers in the
country that may make the AI quality improvements much more fi tting for
certain types of customers than others. Microsoft created an AI program to
learn from Twitter and see if it could create content that people would think
was written by a human. They started it in the United States and had to
shut it down almost immediately because it became so abusive and off ensive.
It mirrored what it saw online. Running the same program in China, where
Twitter is heavily censored, it has performed well and not turned abusive.
The attributes of the product and the “quality” of the product depend on
how relevant the training sample is to that customer.
This is likely to infl uence the adoption rate of the AI technologies in
diff erent places. Again, think of the autonomous cars. Will we gather loads
of information about driving in urban areas and on highways or in Silicon
valley from the early adopters, tailor the product to their needs, but then
fi nd that it does not work as well for dirt roads or rural places or places
without Bay Area weather?
Heterogeneous demand is the hobgoblin of the AI mind. Groups that
diff er most from the training sample will likely be the slowest to adopt the
technology, in part, because it will be the least helpful to them. That may
lead to another manifestation of the digital divide. In this sense, the rise of
AI technologies is likely
to make the problem of income and of geographic
inequality even worse. To the extent that new AI technologies are expensive
and tailored toward the training sample of adopters, it will be like having
lower infl ation and greater consumer surplus going to those groups (for
discussions about diff erences in prices and innovation across income groups
or for online buyers versus offl
ine buyers, see Jaravel [2017] or Goolsbee and
Klenow [2018]).
Government policy will face the potential of divisions along red state/
blue state or high- education/ low- education locations or high- income/
low- income neighborhoods even more than it does today.
11.3 Challenges for Universal Basic Income
as a Response to Job Market Displacement
Now suppose that the arguments above prove wrong. Nothing slows the
speed of AI adoption and there is mass job displacement in a short time.
Public Policy in an AI Economy 313
There has been a rising call among the believers in that scenario for univer-
sal basic income policy. Closely tied to the old Milton Friedman notion of
a negative income tax, the UBI would grant some minimal level of income
to people regardless of employment status as a new form of safety net and
anyone could then work beyond that UBI level to earn more. In the purest
libertarian concept, this UBI would replace the existing collection of safety
net programs. The advantage of the UBI would be that people could survive
in a world with few human jobs and alleviate poverty in a relatively effi
-
cient manner and without destroying all incentives in the private economy.
It seeks to separate the notion of “making a living” from having a job.
There are some small- scale experiments with the UBI in a few countries like
Finland and New Zealand or funded by private individuals in the United
States. There are a number of challenges associated with negative income
taxes and UBIs as a policy solution to widespread AI adoption.
First, if you accept the economists’ basic labor supply model (that people
value leisure and so generally need to be paid to work) then there are likely
to be some sizable number of people who are working only because they
absolutely have to. In a world where AI- induced unemployment is already
high, separating work and income might be an advantage. In a world like the