The Economics of Artificial Intelligence
Page 55
and a safety net that is responsive to need and ensures opportunity—will
aff ect whether we are able to fully reap the benefi ts of AI while also mini-
mizing its potentially disruptive eff ects on the economy and society. And in
the process, such policies could also aff ect productivity growth—including
advances in AI itself.
References
Arntz, Melanie, Terry Gregory, and Ulrich Zierahn. 2016. “The Risk of Automation
for Jobs in OECD Countries: A Comparative Analysis.” OECD Social, Employ-
ment and Migration Working Papers no. 189, Organisation for Economic Co-
operation and Development.
Autor, David. 2014. “Polanyi’s Paradox and the Shape of Employment Growth.”
NBER Working Paper no. 20485, Cambridge, MA.
Byrne, David, John Fernald, and Marshall Reinsdorf. 2016. “Does the United States
Have a Productivity Slowdown or a Measurement Problem?” Brookings Papers on
Economic Activity, Spring 2016. https:// www .brookings .edu/ wp- content/ uploads
/ 2016/ 03/ byrnetextspring16bpea .pdf.
Council of Economic Advisers (CEA). 2016. “The Long- Term Decline in Prime-
Age Male Labor Force Participation.” Report, Executive Offi
ce of the President
of the United States.
Executive Offi
ce of the President (EOP). 2016. “Artifi cial Intelligence, Automation,
and the Economy.” Report.
328 Jason Furman
Frey, Carl, and Michael Osborne. 2013. “The Future of Employment: How Suscep-
tible are Jobs to Computerization.” Unpublished manuscript, Oxford University.
Goldin, Claudia, and Lawrence Katz. 2008. The Race between Education and Tech-
nology. Cambridge, MA: Harvard University Press.
Graetz, Georg, and Guy Michaels. 2015. “Robots at Work.” CEPR Discussion Paper
no. DP10477, Centre for Economic Policy Research.
Murray, Charles. 2006. In Our Hands: A Plan to Replace the Welfare State. Wash-
ington, DC: AEI Press.
Rhodes, Elizabeth, Matt Krisiloff , and Sam Altman. 2016. “Moving Forward on
Basic Income.” Blog, Y Combinator. May 31.
Schmitt, John, Heidi Schierholz, and Lawrence Mishel. 2013. “Don’t Blame the
Robots. Assessing the Job Polarization Explanation of Growing Wage Inequality.”
EPI Working Paper, Economic Policy Institute. https:// www .epi .org/ publication
/ technology- inequality- dont- blame- the- robots//
Stern, Andy, and Lee Kravitz. 2016. Raising the Floor: How a Universal Basic Income Can Renew Our Economy and Rebuild the American Dream. New York: Public-Aff airs.
Syverson, Chad. 2013. “Will History Repeat Itself ? Comments on ‘Is the Informa-
tion Technology Revolution Over?’ ” International Productivity Monitor 25 (2):
37– 40.
———. 2016. “Challenges to Mismeasurement Explanations for the U.S. Productiv-
ity Slowdown.” NBER Working Paper no. 21974, Cambridge, MA.
Western, Bruce, and Jake Rosenfeld. 2011. “Unions, Norms, and the Rise in U.S.
Wage Inequality.” American Sociological Review 76 (4): 513– 37.
13
R&D, Structural Transformation,
and the Distribution of Income
Jeff rey D. Sachs
13.1 Introduction
Oh, for the days of balanced growth. In Solow’s growth model, labor-
augmenting technical change at a constant rate produces long- term growth
in output per capita and wages at the same constant rate. The returns to
capital are stable, as are the factor shares of national income going to labor
and capital. In the heyday of the Solow model, these were viewed by Kaldor
(1957) and others as the stylized facts of long- term economic growth.
These stylized facts have visibly broken down since around the year 2000.
There has been a striking disconnect between the continued growth of labor
productivity (gross domestic product [GDP] per worker) and the stagnation
of compensation per worker, resulting in a discernible decline in the labor
share of income, as shown in fi gure 13.1 for the nonfarm business sector
(Elsby, Hobijn, and S¸ahin 2013; ILO and OECD 2015; Karabarbounis and
Neiman 2013; Koh, Santaeulalia- Llopis, and Zheng 2015). The decline in
labor share is widely, if not universally, attributed to automation—robots
and other smart machines—displacing labor.
There are other possible culprits besides automation, including a conjec-
tured rise in monopoly power, a fall in US union coverage and power, and
the eff ects of global trade on the distribution of income. Of course, several
factors may be at play. My view is that automation—the replacement of
Jeff rey D. Sachs is University Professor, the Quetelet Professor of Sustainable Development, and professor of health policy and management at Columbia University and a research associate of the National Bureau of Economic Research.
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/ c14014.ack.
329
330 Jeff rey D. Sachs
Fig. 13.1 Labor share of GDP at factor cost
Source: Data are from Components of Value Added by Industry, millions of dollars, Bureau of Economic Analysis, release date: November 3, 2016.
Note: The labor share is defi ned as compensation of employees divided by the sum of compensation of employees and operating surplus for the gross domestic product.
human labor by machines and code—is likely to be the most important of
the factors.
Indeed, my argument is that the decline in the labor share via automation
has been occurring well before the year 2000, but that it has been obscured in
the macroeconomic data by off setting structural changes. Balanced growth,
in short, was always a mirage. The diff erence now is that the imbalances are
now showing more vividly, and are likely to intensify.
One reason that unbalanced growth was underemphasized before the year
2000 is that diff erent sectors of the economy were aff ected by automation in
diff erent, and indeed off setting, ways. It is useful, I believe, to disaggregate
GDP into fi ve major sectors:
• goods- producing sectors: agriculture, mining, construction, and manu-
facturing;
• basic business services: utilities, wholesale trade, retail trade, transport,
and warehousing;
• personal services: arts, leisure, food, and accommodations, other per-
sonal;
• professional services: information, fi nance, education, health, manage-
ment, scientifi c and technical, other professional; and
• government services: federal, state, and local.
R&D, Structural Transformation, and the Distribution of Income 331
Table 13.1
Required expertise and workfl ow predictability by sector
Sector
Typical expertise/ education
Typical workfl ow predictability
Goods producing
Low to moderate
High
Basic business services
Moderate
Moderate to high
Personal services
Low to moderate
Low to moderate
Professional services
High
Low
Government
Moderate to high
M
oderate to high
These sectors are diff erentially susceptible to automation. Historically, there
seem to have been two key dimensions to work tasks that determine their
suitability for automation: degree of expertise required and repetitiveness/
predictability of the task (Frey and Osborne 2013; Chui, Manyika, and
Miremadi 2016; McKinsey Global Institute 2017). Tasks requiring high
expertise (e.g., as measured by their educational requirements) and that
have low predictability/ repetitiveness in workfl ow have been less easily auto-
mated. Based on the occupational mix and production processes of the fi ve
sectors, we can place the sectors roughly as seen in table 13.1.
This suggests that the goods- producing sector has been easiest to auto-
mate and professional services the most diffi
cult, with the other sectors
somewhere in the middle, depending on the particular subsectors involved.
As I describe later, artifi cial intelligence (AI) could change the character of
automation in the future, leading to much more automation of high- skill
tasks.
These diff erences in susceptibility to automation show up in the sector
trends in labor share of value added (measured at factor cost) since 1987,
shown in fi gure 13.2.
We see a large drop in the labor share of value added in the goods-
producing sector, from 61.7 percent to 48.9 percent, consistent with the
ease of automation in that sector, contrasted by an increase in the labor
share of value added in professional services and government, consistent
with the relative diffi
culty of automation in those two sectors. Basic business
services also show a modest decline in the labor share, from 66.3 percent
to 60.1 percent. The labor share of value added in personal services was
unchanged, consistent with the relatively low workfl ow predictability of that
sector, making it more diffi
cult to automate.
Figure 13.2 makes clear that in the goods- producing and basic- business-
service sectors, automation has been taking place for decades, but the trends
have been somewhat obscured by the relative lack of automation in the
other sectors, and by the fact that both output and employment have been
shifting from goods production to professional services, that is, from the
broad sectors experiencing the most automation to the ones experiencing
the least automation.
332 Jeff rey D. Sachs
Fig. 13.2 Labor share by sector
Source: Data source fi gure 13.1.
Notes: The labor share by sector is equal to the labor compensation for all subsectors divided by the sum of employee compensation and operating surplus for all subsectors. The sectors are as follows. Goods- producing sector: agriculture, forestry, fi shing and hunting; mining; construction; and manufacturing. Basic business services: utilities, wholesale trade, retail trade, and transportation and warehousing. Professional services: information; fi nance and insurance; professional and business services; and educational services, health care, and social assistance. Personal services: arts, entertainment, recreation, accommodation, and food services. Government includes federal, state, and local government.
Even the signifi cant observed decline in the labor share in the goods-
producing sector understates the extent of structural change in that sector,
since the composition of labor has also been shifting dramatically from
production workers with relatively low levels of schooling to supervisory
workers with higher levels of schooling. This too marks a rise in the share
of capital income in value added, albeit the income earned by human capital
rather than by business fi xed capital.
Figure 13.3 off ers a rough estimate of the overall share of labor income
in the economy accounted for by diff erent levels of educational attainment.
For our purposes, I have grouped the educational attainment into three
bins: low, compromising attainment up to some college including a two- year
associate’s degree; medium, comprising a bachelor’s degree but no advanced
degree; and high, comprising an advanced degree. Using census data on the
mean income and number of workers at these levels of educational attain-
ments, we can fi nd the shares of labor income accruing to diff erent cate-
gories, as shown in fi gure 13.3.
Labor income accruing to workers with less than a bachelor’s degree
R&D, Structural Transformation, and the Distribution of Income 333
Fig. 13.3 Share of earnings by educational attainment
Source: Data are from the United States Census Bureau, table A- 3, “Mean Earnings of Workers 18 Years and Over, by Educational Attainment, Race, Hispanic Origin, and Sex: 1975 to 2015,” (https:// census .gov/ data/ tables/ 2016/ demo/ education- attainment/ cps- detailed- tables
.html). Low education: not a high school graduate; high school graduate; some college or associate’s degree; medium education: bachelor’s degree; high education: advanced degree. Total income by level of educational attainment is the product of the number of workers with earnings and mean earnings.
plummeted from 72.7 percent to 46.1 percent. Workers with a bachelor’s
degree saw their share of labor income doubling from 14.3 percent to
29.6 percent, and workers with an advanced degree also saw their share of
labor income doubling from 12.9 percent to 23.4 percent.
Real mean earnings per worker among these three categories shows a
similar trend in fi gure 13.4. Earnings of low- skilled workers (defi ned here
as all the way up to some college or an associate’s degree) began to stagnate
in the mid- 1970s, and have not risen since then. Mean earnings for workers
with a bachelor’s or advanced degree continued to rise until around the year
2000 and have since been stagnant—or even falling, in real terms, in the case
of those with advanced degrees.
The relative numbers of workers at each educational attainment has
responded to the changing market incentives and to outlays for education by
governments at all levels. As we see in fi gure 13.5, the proportion of all work-
ers at less than a bachelor’s degree declined from 83.4 percent to 64.3 per-
cent, while those with a BA rose from 10.0 to 22.6 percent, and those with
an advanced degree rose from 6.6 to 13.2 percent between 1975 and 2016.
What makes these trends especially important for us, I believe, is that the
ability to automate tasks is likely to increase dramatically with the recent
advances in big data, machine learning, and other forms of artifi cial intel-
ligence. The trends to date—the falling share of labor income, rising share
of earnings fl owing to highly trained workers, and decline of real earnings of
workers who are subject to automation—may soon be felt by a much wider
swath of workers and sectors.
334 Jeff rey D. Sachs
Fig. 13.4 Real mean earnings by education in $1982– 1984
Source: Earnings data from source in fi gure 13.3.
Note: Real mean earnings for each education group are obtained by aggregating total earnings for the educational level, dividing by number of workers with earnings, and defl ating by the Consumer Price Index.
Fig. 13.5 Share of employment by education
Source: Employment data from the source in fi gure 13.3.
In a fundamental sense we are witnessin
g the gradual unfolding of a fun-
damental general purpose technology, digital information, that is at least
as fundamental as the steam engine and electrifi cation. Digital informa-
tion began to unfold with the theoretical breakthroughs of Alan Turing,
John von Neumann, Claude Shannon, and Norbert Weiner in the 1930s and
1940s, and then advanced dramatically with the fi rst mainframe computers
in the 1940s, the invention of the transistor in 1947, the invention of inte-
grated circuits in the late 1950s, and the initiation of Moore’s Law at the end
R&D, Structural Transformation, and the Distribution of Income 335
of the 1950s. Of course, the digital revolution now engages a vast range of
science and technology, including solid- state physics, nanotechnology, fi ber
optics, digital communication, and a startling range of applications across
every domain of science and every sector of the economy.
The rising investments in research and development (R&D) are therefore
a fundamental part of the story and the fundamental driver of structural
transformation. Figure 13.6 shows the national accounts estimates of R&D
annual outlays and the cumulative stock of intellectual property, both as
a share of GDP. Research and development as a share of GDP roughly
doubles from the early 1950s to today, from around 1.3 percent to 2.6 per-
cent. The stock of intellectual property (IP) rises from around 4.5 percent
to 14 percent of GDP. The point is that IP has risen far faster than GDP;
the economy has become far more science intensive.
Rather than the Solow- era stylized facts, I would therefore propose the
following alternative stylized facts:
1. The share of national income accruing to capital rises over time in sec-
tors experiencing automation, especially when capital is measured to include
human capital.
2. The share of national income accruing to low- skilled labor drops while
the share accruing to high- skilled labor rises.
Fig. 13.6 R&D and intellectual property (percent GDP)
Source: The Net Stock of Intellectual Property Products is from the Bureau of Economic Analysis, table 2.1. Current- Cost Net Stock of Private Fixed Assets, Equipment, Structures, and Intellectual Property Products by Type, https:// www .bea .gov/ iTable/ iTable .cfm?reqid=10