The Economics of Artificial Intelligence
Page 53
one we are in now, though, off ering a basic income will likely cause a sizable
drop in labor market participation by low- wage earners. To the extent that
nonparticipation in exactly that segment of the labor force is already viewed
as a problem, the UBI would likely make things worse and risk angering
the broader public.
Second, for a given amount of money to be used on redistribution, a UBI
likely shifts money away from the very poor. To oversimplify, if you have
$50 billion to alleviate poverty, the targeting approach followed in most
countries today might use the $50 billion to help the poorest/ sickest 25 mil-
lion people and give them the equivalent of $25,000 of benefi ts each. With
a broad- based UBI, the same $50 billion would be spread out. It might
involve, say, 100 million people getting $5,000 each. Perhaps a UBI could
change the total taste for redistribution in a society—leaving the most dis-
advantaged people with the same amount and upping the total amount
spent—but for the UBI to not end up more regressive than the current
system necessarily entails greater amounts of public funds.
Third, the conception of the UBI as a replacement for a myriad of other
in-kind transfers and safety net programs forgets the historical origins of
that safety net. Fundamentally, the in-kind safety net exists today because
rich societies are not comfortable with grievously injured people coming
into a hospital but being turned away if they do not have money or letting
kids go hungry because their parents cannot aff ord to feed them, and so on.
Converting to a UBI and abolishing the in-kind safety net will lead to a situa-
tion where some people will blow their UBI money in unsympathetic ways—
gambling, drugs, junk food, Ponzi schemes, whatever. Those people will then
314 Austan Goolsbee
come to emergency rooms or their kids will be hungry and by the rules, they
will be out of luck. That is what their UBI income was supposed to cover.
But the fact that advanced economies evolved an in-kind safety net in order
to avoid this situation makes me think that enforcing “UBI discipline” and
replacing the safety net with a straight transfer would require rather extra-
ordinary changes in the psyche of people in the advanced economies.
11.4 Policy Responses to AI beyond Jobs: Pricing,
Data Property Rights, and Antitrust
Just as the impact of AI goes far beyond just the impact on employment,
the policy response to AI raises all sorts of other considerations, as well.
One is the perennial back- and- forth over the power of buyers versus
the power of sellers in pricing. The same issue arose with the initial rise
of ecommerce—the new online data on customers allowed new forms of
price discrimination and market power but the ease of comparison shopping
reduced search costs and promoted competition (e.g., Brown and Goolsbee
2002). So far, the power of the AI technology seems overwhelmingly to have
been used by sellers. If they can individualize market and price discriminate
with it, margins will likely rise. But consumers will likely push back. They
may fi nd technological solutions to use AI to thwart merchants. But a more
straightforward response might be to follow past practice and start making
various behaviors and practices illegal. This could include restrictions on
consumer privacy and the ways that companies can use customer informa-
tion. It might manifest as an argument over property rights in the sense
of who owns the consumers’ data and what level of consent it requires to
use it, or might involve rules against various types of price discrimination.
Regardless of the form, these issues of pricing and data seem like they will
be a central area of policy in an AI- centric world.
The second thing about an AI economy is that the fi xed- cost/ economies
of scale seem pretty signifi cant, and in many cases there are also often net-
work externalities and switching costs on the demand side of these indus-
tries. All of these seem to portend the possibility of many industries having a
winner- take- all market structure or the continued rise of “platform” com-
petition rather than conventional competition. If so, the rise of AI is likely
to usher in a renewed emphasis on antitrust policy in much the same way
the original Gilded Age consolidation of industry did before.
11.5 Conclusion: Will Robots Take Over Policy, Too?
The organizers of the volume also asked us to consider whether AI will
enhance or even replace the jobs of policy makers—whether improvements
in machine learning and AI could be used on the policy- making process
itself. Personally, I do not think so because the most important policy mat-
Public Policy in an AI Economy 315
ters are at their heart not issues of prediction. The technology may improve
our ability to predict responses, but it does not help us balance interests or
engage in politics. We already know, for example, a great deal about the fi s-
cal implications for social security of the aging population. Artifi cial intel-
ligence might improve our ability to predict revenue outcomes of various
policy options, say. That has not been the problem with addressing social
security. It has always been about choosing between options and making
value judgments. The kinds of problems that AI helps with are those where
large amounts of past data to inform the decision. Conditions with small
samples or where the conditions are very diff erent than in the past will
be much less machine learnable. For small bore issues, AI may improve
policy accuracy—what conditions should cause regulators to raise their
estimated probability that a bank’s loans will start to default, for example.
For bigger issues, though, like whether the Federal Reserve should raise
interest rates or whether we should cut high- income people’s taxes—I have
my doubts about what AI can contribute.
It is also sure to increase the attention paid to business practices of large
AI platforms—their pricing, their use of personal data on customers, their
behavior toward competitors, and the continuing consolidation of market
power. Each of these is likely to become a major policy battleground of the
future. For the time being, though, the job of policy makers themselves seem
relatively safe . . . for now.
References
Autor, David. 2015. “Why Are There Still So Many Jobs? The History and Future of
Workplace Automation.” Journal of Economic Perspectives 29 (3): 3– 30.
Autor, David, and A. M. Salomons. 2018. “Is Automation Labor- Displacing?
Productivity Growth, Employment, and the Labor Share.” Brookings Papers
on Economic Activity 2018, Spring. https:// www .brookings .edu/ bpea- articles
/ is- automation- labor- displacing- productivity- growth- employment- and- the
- labor- share/.
Bresnahan, Timothy, and Robert Gordon, eds. 1997. The Economics of New Goods.
Chicago: University of Chicago Press.
Brown, Jeff rey, and Austan D. Goolsbee. 2002. “Does the Internet Make Markets
More Competitive? Evidence from the Life Insurance Industry.” Journal of P
o-
litical Economy 110 (3): 481– 507.
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.
Goolsbee, Austan D., and Peter J. Klenow. 2006. “Valuing Consumer Goods by the
Time Spent Using Them: An Application to the Internet.” American Economic
Review, Papers and Proceedings 96 (2): 108– 13.
———. 2018. “Internet Rising, Prices Falling: Measuring Infl ation in a World of
E-Commerce.” American Economic Review, Papers and Proceedings 108 (5): 488– 92.
316 Austan Goolsbee
Gordon, Robert. 2016. The Rise and Fall of American Growth: The US Standard of
Living since the Civil War. Princeton, NJ: Princeton University Press.
McKinsey Global Institute. 2017. Jobs Lost, Jobs Gained: Workforce Transitions in a Time of Automation. December. McKinsey & Co. Accessed Apr. 26, 2018. https://
www .mckinsey .com/ ~/ media/ McKinsey/ Global%20Themes/ Future%20of%20
Organizations/ What%20the%20future%20of%20work%20will%20mean%20
for%20jobs%20skills%20and%20wages/ MGI- Jobs- Lost- Jobs- Gained- Report
- December- 6– 2017.ashx.
Jaravel, Xavier. 2017. “The Unequal Gains from Product Innovations: Evidence
from the US Retail Sector.” Unpublished manuscript, London School of Eco-
nomics. April.
Mokyr, Joel. 2014. “Secular Stagnation? Not in Your Life.” In Secular Stagnation:
Facts Causes and Cures, edited by Coen Teulings and Richard Baldwin, 83– 89.
London: CEPR Press.
Varian, Hal. 2013. “The Value of the Internet, Now and in the Future.” Economist, Mar. 10. Accessed Apr. 26, 2018. https:// www .economist .com/ blogs/ freeexchange
/ 2013/ 03/ technology- 1.
12
Should We Be Reassured If
Automation in the Future Looks
Like Automation in the Past?
Jason Furman
Much of the debate about the economic impact of artifi cial intelligence
(AI) centers on the question of whether this time will be diff erent. Some
optimists argue that AI is no diff erent than technologies that came before
it and that centuries of fears that machines will replace human labor have
proven unfounded, with machines instead creating previously unimagined
jobs and raising incomes. Others argue that AI is diff erent—by replacing
cognitive tasks, it could render much of human employment redundant,
leading to mass unemployment in the eyes of the pessimists or historically
unparalleled freedom for leisure in the eyes of the optimists.
The history of automation—and how the US economy has handled it
over the last several decades—suggests that even if AI is similar to pre-
vious waves of automation, that should not be entirely comforting since
technological advances in recent decades have brought tremendous benefi ts
but have also contributed to increasing inequality and falling labor force
participation. This outcome, however, is not inevitable because the eff ects
of technological change on the workforce are mediated by a wide set of
institutions, and as such, policy choices will have a major impact on actual
outcomes. Artifi cial intelligence does not call for a completely new paradigm
for economic policy—for example, as advocated by proponents of replac-
ing the existing social safety net with a universal basic income (UBI)—but
instead reinforces many of the steps that could already be justifi ed by the
goal of making sure that growth is shared more broadly.
To date, in fact, the problem we have faced is not too much automation
Jason Furman is professor of the practice of economic policy at Harvard Kennedy School and a nonresident senior fellow at the Peterson Institute for International Economics.
For acknowledgments, sources of research support, and disclosure of the author’s material fi nancial relationships, if any, please see http:// www .nber .org/ chapters/ c14031.ack.
317
318 Jason Furman
but too little automation—the issue I will address before considering some
of the potentially harmful side eff ects that a faster pace of innovation can
have for inequality and labor force participation. In the course of this discus-
sion I will address the extent to which policy can advance AI while ensuring
that more people share in the benefi ts of it, two goals that are ultimately
complementary.
12.1 The
Benefi t of More Artifi cial Intelligence
Technologists see transformative change all around us but economists
are a more sour bunch, focusing on productivity statistics that show that
we are adding very little to output per hour. Measured productivity growth
has slowed in thirty- fi ve of thirty- six advanced economies, slowing from a
2.7 percent average annual growth rate from 1996 to 2006 to a 1.0 percent
average annual growth rate from 2006 to 2016—with the slowdown in the
G7 economies shown in fi gure 12.1.
There are many reasons to believe that the offi
cial statistics fail to capture
the full range of productivity improvements, so the 1.0 percent estimate
likely understates productivity growth from 2006 to 2016. But so, too, does
the 2.7 percent fi gure understate productivity growth from 1996 to 2006, a
period that witnessed the de facto invention of the World Wide Web and
its associated uses for search, ecommerce, email, and much more—not to
mention the widespread adoption of cellphones and invention of mobile
Fig. 12.1 Labor productivity growth, G- 7 countries
Source: The Conference Board, Total Economy Database; author’s calculations.
If Automation in the Future Looks Like Automation in the Past 319
Fig. 12.2 Estimated worldwide annual supply of industrial robots, 2006– 2016
Source: International Federation of Robotics, World Robotics (2016, 2017).
email. Recent research has confi rmed that there is little reason to doubt the
magnitude of the reduction in productivity growth, including pointing out
that the slowdown has also occurred in well- measured industries (Byrne,
Fernald, and Reinsdorf 2016; Syverson 2016).
This may seem counterintuitive given all the excitement around new inno-
vations—including in robotics, AI, and automation more generally—but as
exciting as these innovations may be, they still represent only a tiny fraction
of our lives when compared to other sectors of the economy like housing,
retail, education, and health—and, at least to date, the improvements they
are making in these sectors are not dramatically diff erent than the improve-
ments we saw in previous eras of the economy.
That said, the technology sector of our economy is making important
contributions to productivity growth. A 2015 study of robots in seventeen
countries found that they added an estimated 0.4 percentage point on average
to those countries’ annual gross domestic product (GDP) growth between
1993 and 2007, accounting for a bit more than one- tenth of those countries’
overall GDP growth during that time (Graetz and Michaels 2015). More-
over, since 2010, worldwide shipments of industrial robots have increased
dramatically, as shown in fi gure 12.2, potentially signaling even more pro-
ductivity growth in the future.
Relatedly, there has been dramatic progress in recent years in AI and its
application in a diverse set of areas. For example, companies are using AI
to analyze online customer transactions in order to detect and stop fraud,
and, similarly, social networking sites are using it to detect when an account
may have been hijacked. Thanks to AI, web search applications are now
more accurate—for example, by correcting for manual entry error—thereby
reducing costs associated with search. In radiology, where doctors must be
320 Jason Furman
able to examine radiological images for irregularities, AI’s superior image
processing techniques may soon be able to provide more accurate image
analysis, expanding the potential for earlier detection of harmful abnormali-
ties and reducing false positives, ultimately leading to better care. Artifi cial
intelligence is also making inroads in the public sector as well. For example,
predictive analytics has great potential to improve criminal justice proce-
dures, although it must be used responsibly to avoid bias.
However, while AI research has been underway for decades, recent
advances are still very new, and, as a result, AI has not had a large macro-
economic impact, at least not yet. The most recent major progress in AI has
been in deep learning, a powerful method but one that must be applied in
a customized way for each application. Even though we have not made as
much progress recently on other areas of AI, such as logical reasoning, the
advancements in deep- learning techniques may ultimately act as at least a
partial substitute for these other areas.
While AI has an advantage over humans in many respects, humans still
maintain a substantial advantage over AI for tasks that involve social intel-
ligence, creativity, and general intelligence. For example, AI today can do
decent translations but cannot come close to what a human can do with his
or her knowledge of both languages, social and cultural context, and sense
of the author’s argument, emotional states, and intentions. As it stands,
even the most popular machine translator still fails to reach the accuracy
of a human translator.
It is possible that major new inventions like electricity have manifested