Content uploaded by David Autor
Author content
All content in this area was uploaded by David Autor on Nov 12, 2015
Content may be subject to copyright.
Journal of Economic Perspectives—Volume 29, Number 3—Summer 2015—Pages 3–30
T
here have been periodic warnings in the last two centuries that automation
and new technology were going to wipe out large numbers of middle class
jobs. The best-known early example is the Luddite movement of the early
19thcentury, in which a group of English textile artisans protested the automation
of textile production by seeking to destroy some of the machines. A lesser-known
but more recent example is the concern over “The Automation Jobless,” as they
were called in the title of a TIME magazine story of February 24, 1961:
The number of jobs lost to more efficient machines is only part of the prob-
lem. What worries many job experts more is that automation may prevent
the economy from creating enough new jobs. . . . Throughout industry, the
trend has been to bigger production with a smaller work force. . . . Many of
the losses in factory jobs have been countered by an increase in the service
industries or in office jobs. But automation is beginning to move in and elimi-
nate office jobs too. . . . In the past, new industries hired far more people
than those they put out of business. But this is not true of many of today’s
new industries. . . . Today’s new industries have comparatively few jobs for
the unskilled or semiskilled, just the class of workers whose jobs are being
eliminated by automation.
Concerns over automation and joblessness during the 1950s and early 1960s
were strong enough that in 1964, President Lyndon B. Johnson empaneled a
Why Are There Still So Many Jobs?
The History and Future of Workplace
Automation
†
■ David H. Autor is Professor of Economics, Massachusetts Institute of Technology, Cambridge,
Massachusetts. From 2009 to 2014, he was Editor of the Journal of Economic Perspectives.
†
To access the Data Appendix and disclosure statement, visit
http://dx.doi.org/10.1257/jep.29.3.3 doi=10.1257/jep.29.3.3
David H. Autor
4 Journal of Economic Perspectives
“Blue-Ribbon National Commission on Technology, Automation, and Economic
Progress” to confront the productivity problem of that period—specifically, the
problem that productivity was rising so fast it might outstrip demand for labor.
The commission ultimately concluded that automation did not threaten employ-
ment: “Thus technological change (along with other forms of economic change) is
an important determinant of the precise places, industries, and people affected by
unemployment. But the general level of demand for goods and services is by far the
most important factor determining how many are affected, how long they stay unem-
ployed, and how hard it is for new entrants to the labor market to find jobs. The
basic fact is that technology eliminates jobs, not work” (Bowen 1966, p. 9). However,
the Commission took the reality of technological disruption as severe enough that
it recommended, as one newspaper (The Herald Post 1966) reported, “a guaranteed
minimum income for each family; using the government as the employer of last
resort for the hard core jobless; two years of free education in either community
or vocational colleges; a fully administered federal employment service, and indi-
vidual Federal Reserve Bank sponsorship in area economic development free from
the Fed’s national headquarters.”
Such concerns have recently regained prominence. In their widely discussed book
The Second Machine Age, MIT scholars Erik Brynjolfsson and Andrew McAfee (2014,
p. 11) offer an unsettling picture of the likely effects of automation onemployment:
Rapid and accelerating digitization is likely to bring economic rather than
environmental disruption, stemming from the fact that as computers get more
powerful, companies have less need for some kinds of workers. Technological
progress is going to leave behind some people, perhaps even a lot of people,
as it races ahead. As we’ll demonstrate, there’s never been a better time to be a
worker with special skills or the right education, because these people can use
technology to create and capture value. However, there’s never been a worse
time to be a worker with only ‘ordinary’ skills and abilities to offer, because
computers, robots, and other digital technologies are acquiring these skills
and abilities at an extraordinary rate.
Clearly, the past two centuries of automation and technological progress have
not made human labor obsolete: the employment‐to‐population ratio rose during
the 20thcentury even as women moved from home to market; and although the
unemployment rate fluctuates cyclically, there is no apparent long-run increase. But
those concerned about automation and employment are quick to point out that
past interactions between automation and employment cannot settle arguments
about how these elements might interact in the future: in particular, the emergence
of greatly improved computing power, artificial intelligence, and robotics raises the
possibility of replacing labor on a scale not previously observed. There is no funda-
mental economic law that guarantees every adult will be able to earn a living solely
on the basis of sound mind and good character. Whatever the future holds, the
present clearly offers a resurgence of automation anxiety (Akst 2013).
David H. Autor 5
In this essay, I begin by identifying the reasons that automation has not wiped
out a majority of jobs over the decades and centuries. Automation does indeed
substitute for labor—as it is typically intended to do. However, automation also
complements labor, raises output in ways that lead to higher demand for labor, and
interacts with adjustments in labor supply. Indeed, a key observation of the paper
is that journalists and even expert commentators tend to overstate the extent of
machine substitution for human labor and ignore the strong complementarities
between automation and labor that increase productivity, raise earnings, and
augment demand for labor.
Changes in technology do alter the types of jobs available and what those jobs
pay. In the last few decades, one noticeable change has been “polarization” of the
labor market, in which wage gains went disproportionately to those at the top and
at the bottom of the income and skill distribution, not to those in the middle. I will
offer some evidence on this phenomenon. However, I will also argue that this polar-
ization is unlikely to continue very far into the foreseeable future.
The final section of this paper reflects on how recent and future advances in arti-
ficial intelligence and robotics should shape our thinking about the likely trajectory
of occupational change and employment growth. I argue that the interplay between
machine and human comparative advantage allows computers to substitute for workers
in performing routine, codifiable tasks while amplifying the comparative advantage of
workers in supplying problem-solving skills, adaptability, and creativity. The frontier
of automation is rapidly advancing, and the challenges to substituting machines for
workers in tasks requiring flexibility, judgment, and common sense remain immense.
In many cases, machines both substitute for and complement human labor. Focusing
only on what is lost misses a central economic mechanism by which automation affects
the demand for labor: raising the value of the tasks that workers uniquely supply.
How Automation and Employment Interact
In 1900, 41 percent of the US workforce was employed in agriculture; by
2000, that share had fallen to 2 percent (Autor 2014), mostly due to a wide range
of technologies including automated machinery. The mass-produced automo-
bile drastically reduced demand for many equestrian occupations, including
blacksmiths and stable hands. Successive waves of earth-moving equipment and
powered tools displaced manual labor from construction. In more recent years,
when a computer processes a company’s payroll, alphabetizes a list of names, or
tabulates the age distribution of residents in each Census enumeration district,
it is replacing a task that a human would have done in a previous era. Broadly
speaking, many—perhaps most—workplace technologies are designed to save
labor. Whether the technology is tractors, assembly lines, or spreadsheets,
the first-order goal is to substitute mechanical power for human musculature,
machine-consistency for human handiwork, and digital calculation for slow and
error-prone “wetware.”
6 Journal of Economic Perspectives
Given that these technologies demonstrably succeed in their labor saving
objective and, moreover, that we invent many more labor-saving technologies all
the time, should we not be somewhat surprised that technological change hasn’t
already wiped out employment for the vast majority of workers? Why doesn’t auto-
mation necessarily reduce aggregate employment, even as it demonstrably reduces
labor requirements per unit of output produced?
These questions underline an economic reality that is as fundamental as it is over-
looked: tasks that cannot be substituted by automation are generally complemented
by it. Most work processes draw upon a multifaceted set of inputs: labor and capital;
brains and brawn; creativity and rote repetition; technical mastery and intuitive judg-
ment; perspiration and inspiration; adherence to rules and judicious application of
discretion. Typically, these inputs each play essential roles; that is, improvements in
one do not obviate the need for the other. If so, productivity improvements in one set
of tasks almost necessarily increase the economic value of the remaining tasks.
An iconic representation of this idea is found in the O-ring production function
studied by Kremer (1993).
1
In the O-ring model, failure of any one step in the chain
of production leads the entire production process to fail. Conversely, improvements
in the reliability of any given link increase the value of improvements in all of the
others. Intuitively, if n − 1 links in the chain are reasonably likely to fail, the fact
that link n is somewhat unreliable is of little consequence. If the other n − 1 links
are made reliable, then the value of making link n more reliable as well rises. Analo-
gously, when automation or computerization makes some steps in a work process
more reliable, cheaper, or faster, this increases the value of the remaining human
links in the production chain.
As a contemporary example, consider the surprising complementarities between
information technology and employment in banking, specifically the experience with
automated teller machines (ATMs) and bank tellers documented by Bessen (2015).
ATMs were introduced in the 1970s, and their numbers in the US economy quadrupled
from approximately 100,000 to 400,000 between 1995 and 2010. One might naturally
assume that these machines had all but eliminated bank tellers in that interval. But
US bank teller employment actually rose modestly from 500,000 to approximately
550,000 over the 30-year period from 1980 to 2010 (although given the growth in the
labor force in this time interval, these numbers do imply that bank tellers declined
as a share of overall US employment). With the growth of ATMs, what are all of these
tellers doing? Bessen observes that two forces worked in opposite directions. First, by
reducing the cost of operating a bank branch, ATMs indirectly increased the demand
for tellers: the number of tellers per branch fell by more than a third between 1988
and 2004, but the number of urban bank branches (also encouraged by a wave of
1
The name of the O-ring production function refers to the 1986 accident of Space Shuttle Challenger,
which exploded and crashed back to earth less than two minutes after takeoff, killing its seven crew
members. The proximate cause of the Challenger crash was an inexpensive and seemingly inconsequen-
tial rubber O-ring seal in one of its booster rockets that failed after hardening and cracking during the
icy Florida weather on the night before takeoff.
Why Are There Still So Many Jobs? 7
bank deregulation allowing more branches) rose by more than 40 percent. Second,
as the routine cash-handling tasks of bank tellers receded, information technology
also enabled a broader range of bank personnel to become involved in “relationship
banking.” Increasingly, banks recognized the value of tellers enabled by information
technology, not primarily as checkout clerks, but as salespersons, forging relation-
ships with customers and introducing them to additional bank services like credit
cards, loans, and investment products.
This example should not be taken as paradigmatic; technological change is not
necessarily employment-increasing or Pareto-improving. Three main factors can
mitigate or augment its impacts. First, workers are more likely to benefit directly
from automation if they supply tasks that are complemented by automation, but
not if they primarily (or exclusively) supply tasks that are substituted. A construc-
tion worker who is expert with a shovel but cannot drive an excavator will generally
experience falling wages as automation advances. Similarly, a bank teller who can
tally currency but cannot provide “relationship banking” is unlikely to fare well at a
modern bank.
Second, the elasticity of labor supply can mitigate wage gains. If the complemen-
tary tasks that construction workers or relationship bankers supply are abundantly
available elsewhere in the economy, then it is plausible that a flood of new workers
will temper any wage gains that would emanate from complementarities between
automation and human labor input. While these kinds of supply effects will prob-
ably not offset productivity-driven wage gains fully, one can find extreme examples:
Hsieh and Moretti (2003) document that new entry into the real estate broker occu-
pation in response to rising house prices fully offsets average wage gains that would
otherwise have occurred.
Third, the output elasticity of demand combined with income elasticity of
demand can either dampen or amplify the gains from automation. In the case
of agricultural products over the long run, spectacular productivity improvements
have been accompanied by declines in the share of household income spent on food.
In other cases, such as the health care sector, improvements in technology have led
to ever-larger shares of income being spent on health. Even if the elasticity of final
demand for a given sector is below unity—meaning that the sector shrinks as produc-
tivity rises—this does not imply that aggregate demand falls as technology advances;
clearly, the surplus income can be spent elsewhere. As passenger cars displaced eques-
trian travel and the myriad occupations that supported it in the 1920s, the roadside
motel and fast food industries rose up to serve the “motoring public” ( Jackson 1993).
Rising income may also spur demand for activities that have nothing to do with the
technological vanguard. Production of restaurant meals, cleaning services, haircare,
and personal fitness is neither strongly complemented nor substituted by current
technologies; these sectors are “technologically lagging” in Baumol’s (1967) phrase.
But demand for these goods appears strongly income-elastic, so that rising produc-
tivity in technologically leading sectors may boost employment nevertheless in these
activities. Ultimately, this outcome requires that the elasticity of substitution between
leading and lagging sectors is less than or equal to unity (Autor and Dorn 2013).
8 Journal of Economic Perspectives
Over the very long run, gains in productivity have not led to a shortfall of
demand for goods and services: instead, household consumption has largely kept
pace with household incomes. We know this because the share of the population
engaged in paid employment has generally risen over (at least) the past century
despite vast improvements in material standards of living. An average US worker in
2015 wishing to live at the income level of an average worker in 1915 could roughly
achieve this goal by working about 17 weeks per year.
2
Most citizens would not
consider this tradeoff between hours and income desirable, however, suggesting
that consumption demands have risen along with productivity. Of course, citizens
in high-income countries work fewer annual hours, take more vacations, and retire
earlier (relative to death) than a century ago—implying that they choose to spend
part of their rising incomes on increased leisure. This is clearly good news on many
fronts, but does it also imply that consumption demands are approaching satia-
tion? I think not. In high-income countries, consumption and leisure appear to be
complements; citizens spend much of their leisure time consuming—shopping,
traveling, dining, and, less pleasantly, obtaining medical care.
3
What about the Marxian concern that automation will immiserate workers by
obviating the demand for labor? In simple economic models, this outcome cannot
really occur because capital is owned by the economic agents who are presumably
also the workers; but, alternatively, the returns could accrue to a narrow subset of
agents. Sachs and Kotlikoff (2012) and Sachs, Benzell, and LaGarda (2015) explore
multigenerational economic environments in which a burst of robotic productivity
can enrich one generation of capital owners at the expense of future generations.
These later generations suffer because the fruits of the productivity surge are
consumed by the old, while the young face diminished demand for their labor and,
in some cases, also experience credit constraints that inhibit their human capital
investments. In these models, the fundamental threat is not technology per se but
misgovernance; an appropriate capital tax will render the technological advance
broadly welfare-improving, as these papers stress. Thus, a key takeaway is that rapid
automation may create distributional challenges that invite a broad policy response,
a point to which I will return.
2
Douglas (1930; reproduced in US Bureau of the Census 1949) reports average annual earnings across
all sectors in 1915 at $633. Inflating this to 2015 dollars using the US Bureau of Labor Statistics historical
Consumer Price Index calculator yields a current dollar equivalent of $14,711. The BLS employment
report from April 2015 reports mean weekly private nonfarm earnings of $858. Thus, it would take
17 weeks of work at the average US weekly wage to earn a full-time annual 1915 income.
3
This outcome is a modern version of the “coal paradox” posed by William Stanley Jevons in his 1865
book The Coal Question. Jevons argued that as we became more efficient in mining coal, we would use
more of it, not less. Modern environmental economists term this idea the “rebound effect.” In this discus-
sion, the broad parallel is that greater efficiency of production of all goods and services means that we
consume more of them, not the same or less.
David H. Autor 9
Polarization in the US Labor Market
Even if automation does not reduce the quantity of jobs, it may greatly affect the
qualities of jobs available. For the three decades or so from the end of World WarII
and up through the late 1970s, the US experienced rapid automation and tech-
nological change—inspiring, for example, the TIME magazine story in 1961 and
Lyndon Johnson’s 1964 National Commission mentioned earlier. While it’s diffi-
cult to paint an accurate picture of occupational change over a large time interval,
Figure 1, which draws from Katz and Margo (2014), provides a high-level overview
by depicting the average change per decade in employment for seven broad occu-
pational categories, ranked from lowest to highest paid, for two periods: 1940–1980
and 1980–2010. In the first four decades after World WarII, the thrust of occupa-
tional change skewed strongly away from physically demanding, dangerous, and
Figure 1
Average Change per Decade in US Occupational Employment Shares for
Two Periods: 1940–1980 and 1980–2010
Source: Based on Katz and Margo (2014), table 1.6, panel A, which is based upon the 1920 through 2000
Census of population IPUMS and 2010 American Community Survey.
Notes: Observed long changes in US occupational employment shares over 1940–1980 and 1980–2010
are scaled by the number of intervening decades to yield average change per decade. Occupations are
classified into occupational groups based on 1950 occupation codes using the consistent coding of
occupations in all years into 1950 codes (the OCC1950 variable) in the IPUMS. Additional details are
found in Katz and Margo (2014, p. 46).
−4.0
−3.5
−3.0
−2.5
−2.0
−1.5
−1.0
−0.5
0.0
0.5
1.0
1.5
2.0
2.5
3.0
Agricultural
occupations
Service
occupations
Operative/
laborers
Clerical/salesSkilled blue
collar (craft)
Managers Professionals/
technicals
1940–1980
1980–2010
10 Journal of Economic Perspectives
menial work and towards skilled blue- and white-collar work. Agricultural employ-
ment declined by almost 4 percentage points per decade. Professional, technical,
and managerial employment—the highest skill categories—grew by 3 percentage
points per decade (2.5 for the professionals and technicians plus 0.5 for the
managers). And among the vast middle group of workers between agriculture (at
the bottom) and professional, technical, and managerial (the three groups at the
top), service and skilled blue-collar occupations were stable, clerical/sales occupa-
tions rose, and operative and laborer occupations fell sharply.
Thus, physically demanding, repetitive, dangerous, and cognitively monot-
onous work was receding, ushered out by extraordinary productivity gains in
agriculture. Rising consumer affluence spurred demand for manufactured goods
and leisure complements. Growth of technologically intensive corporations,
health care services, and higher education created employment for credentialed
professionals and a cadre of supporting clerical, administrative, and sales workers.
Though automation was clearly reducing labor demand across a large swath of
occupations, it is easy to see why overall job prospects appeared broadly favorable
during thisperiod.
But after the late 1970s, these favorable winds slowed and in some cases
reversed. While jobs at the top of the skill ladder—professional, technical, and
managerial occupations—grew even more rapidly between 1980 and 2010 than
in the four decades prior, positive occupational shifts outside of these catego-
ries mostly halted. Skilled blue-collar occupations shrank rapidly and clerical and
sales occupations—the vulnerable “production jobs” of the information age—
sharply reversed course. While physically demanding operative and laborer jobs
continued to atrophy, low-paid personal services began absorbing an increasing
share of noncollege labor. By this time, the vast movement away from agricultural
work had already played out.
Many forces distinguish the labor markets of these two epochs of 1940–1980
and 1980–2010: a partial list would include changes in the relative supply of college
and noncollege labor, rising trade penetration, offshoring, and globalization of
production chains, declines in labor union penetration, the changing “bite” of the
minimum wage, and certain shifts in tax policy. Of course, many of these factors
combine and interact as well such that attributing changes to a single cause would be
foolish. However, my focus here is on the effects of technological change, and espe-
cially information technology, on employment and occupations (and later wages).
To understand the role that information technology has played (and may play), it
is useful to start from first principles: What do computers do? And how does their
widespread adoption change what workers do?
Fundamentally, computers follow procedures meticulously laid out by program-
mers. The typical pattern has been that for a computer to accomplish a task, a
programmer must first fully understand the sequence of steps required to perform
that task, and then must write a program that, in effect, causes the machine to simu-
late these steps precisely. (The field of machine learning, discussed below, provides
an interesting exception to this process.) When a computer processes a company’s
Why Are There Still So Many Jobs? 11
payroll, alphabetizes a list of names, or tabulates the age distribution of residents
in each Census enumeration district, it is “simulating” a work process that would, in
a previous era, have been done by humans using nearly identical procedures. The
principle of computer simulation of workplace tasks has not fundamentally changed
since the dawn of the computer era—but its cost has. An ingenious 2007 paper
by William Nordhaus estimates that the cost of performing a standardized set of
computations has fallen by at least 1.7 trillion-fold since the manual computing era,
with most of that decline occurring since 1980. Thus, firms have strong economic
incentives to substitute ever-cheaper computing power for relatively expensive
human labor. What are the effects?
One first-order effect is, of course, substitution. As the price of computing
power has fallen, computers and their robot cousins have increasingly displaced
workers in accomplishing explicit, codifiable tasks. In Autor, Levy, and Murnane
(2003), my coauthors and I label these activities as “routine tasks,” not because
they are mundane, but because they can be fully codified and hence automated
(see Levy and Murnane 2004 for many examples). Routine tasks are characteristic
of many middle-skilled cognitive and manual activities: for example, the math-
ematical calculations involved in simple bookkeeping; the retrieving, sorting,
and storing of structured information typical of clerical work; and the precise
executing of a repetitive physical operation in an unchanging environment as in
repetitive production tasks. Because core tasks of these occupations follow precise,
well-understood procedures, they are increasingly codified in computer software
and performed by machines. This force has led to a substantial decline in employ-
ment in clerical, administrative support, and to a lesser degree, in production and
operative employment.
But the scope for this kind of substitution is bounded because there are many
tasks that people understand tacitly and accomplish effortlessly but for which
neither computer programmers nor anyone else can enunciate the explicit “rules”
or procedures. I have referred to this constraint as Polanyi’s paradox, named after
the economist, philosopher, and chemist who observed in 1966, “We know more
than we can tell” (Polanyi 1966; Autor 2015). When we break an egg over the
edge of a mixing bowl, identify a distinct species of birds based on a fleeting
glimpse, write a persuasive paragraph, or develop a hypothesis to explain a poorly
understood phenomenon, we are engaging in tasks that we only tacitly under-
stand how to perform. Following Polanyi’s observation, the tasks that have proved
most vexing to automate are those demanding flexibility, judgment, and common
sense—skills that we understand only tacitly.
4
Polanyi’s paradox also suggests why high-level reasoning is straightforward to
computerize and certain sensorimotor skills are not. High-level reasoning uses a set
4
Computer scientists often refer to this phenomenon as Moravec’s paradox, after Moravec (1988) who
wrote, “[I]t is comparatively easy to make computers exhibit adult level performance on intelligence tests
or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to
perception and mobility.”
12 Journal of Economic Perspectives
of formal logical tools that were developed specifically to address formal problems:
for example, counting, mathematics, logical deduction, and encoding quantita-
tive relationships. In contrast, sensorimotor skills, physical flexibility, common
sense, judgment, intuition, creativity, and spoken language are capabilities that the
human species evolved, rather than developed. Formalizing these skills requires
reverse-engineering a set of activities that we normally accomplish using only tacit
understanding. Hoffman and Furcht (2014) discuss the challenge that Polanyi’s
paradox poses for scientific innovation more broadly.
If computers largely substitute for routine tasks, how do we characterize the
nonroutine tasks for which they do not substitute? In Autor, Levy, and Murnane
(2003), we distinguish two broad sets of tasks that have proven stubbornly challenging
to computerize. One category includes tasks that require problem-solving capabili-
ties, intuition, creativity, and persuasion. These tasks, which we term “abstract,” are
characteristic of professional, technical, and managerial occupations. They employ
workers with high levels of education and analytical capability, and they place a
premium on inductive reasoning, communications ability, and expert mastery. The
second broad category includes tasks requiring situational adaptability, visual and
language recognition, and in-person interactions—which we call “manual” tasks.
Manual tasks are characteristic of food preparation and serving jobs, cleaning and
janitorial work, grounds cleaning and maintenance, in-person health assistance by
home health aides, and numerous jobs in security and protective services. These
jobs tend to employ workers who are physically adept and, in some cases, able to
communicate fluently in spoken language. While these activities are not highly
skilled by the standards of the US labor market, they present daunting challenges
for automation. Equally noteworthy, many outputs of these manual task jobs (hair-
cuts, fresh meals, housecleaning) must be produced and performed largely on-site
or in person (at least for now), and hence these tasks are not subject to outsourcing.
The potential supply of workers who can perform these jobs is very large.
Because jobs that are intensive in either abstract or manual tasks are gener-
ally found at opposite ends of the occupational skill spectrum—in professional,
managerial, and technical occupations on the one hand, and in service and laborer
occupations on the other—this reasoning implies that computerization of “routine”
job tasks may lead to the simultaneous growth of high-education, high-wage jobs at
one end and low-education, low-wage jobs at the other end, both at the expense
of middle-wage, middle education jobs—a phenomenon that Goos and Manning
(2003) called “job polarization.” A large body of US and international evidence
confirms the presence of employment polarization at the level of industries, locali-
ties, and national labor markets (Autor, Katz, and Kearney 2006, 2008; Goos and
Manning 2007; Autor and Dorn 2013; Michaels, Natraj, and Van Reenen 2014;
Goos, Manning, and Salomons 2014; Graetz and Michaels 2015; Autor, Dorn, and
Hanson 2015).
5
5
Mishel, Shierholz, and Schmitt (2013) offer an extended, and for the most part extremely careful,
critique of the literature on technological change, employment, and wage inequality. Their paper argues
David H. Autor 13
Figure 2 illustrates this pattern for the United States by plotting percentage
point changes in employment by decade for the years 1979–2012 for ten major
occupational groups encompassing all of US nonagricultural employment. (More
that the growth of low-wage service employment does not commence in the United States until the
2000s, a finding that is at odds with all other work using contemporary occupation codes of which I am
aware (including the Bureau of Labor Statistic’s own tabulations of Occupational Employment Statistics
data for this time period provided in Alpert and Auyer 2003, table 1). At a methodological level, work in
this area always requires adjustments and judgment calls in comparing occupational data across Census
years, but the adjustments that Mishel et al. apply to the data generate occupational patterns that appear
anomalous. Substantively, I believe the main issue is not whether employment polarization has occurred—
on this, the evidence appears unambiguous—but the extent to which these occupational employment
shifts are helpful for understanding wage polarization or wage inequality more broadly.
Figure 2
Change in Employment by Major Occupational Category, 1979–2012
(the y-axis plots 100 times log changes in employment, which is nearly equivalent to
percentage points for small changes)
Sources: Author using data from the 1980, 1990, and 2000 Census IPUMS files, American Community Survey
combined file 2006–2008, and American Community Survey 2012. The sample includes the working-age
(16–64) civilian noninstitutionalized population. Employment is measured as full-time equivalent workers.
Notes: Figure 2 plots percentage point changes in employment (more precisely, the figure plots 100 times
log changes in employment, which is close to equivalent to percentage points for small changes) by decade
for the years 1979–2012 for ten major occupational groups encompassing all of US nonagricultural
employment. Agricultural occupations comprise no more than 2.2 percent of employment in this time
interval, so this omission has a negligible effect.
−20
−10
0
10
20
30
40
100 × log Change in Employment
Personal care
Food/cleaning service
Protective service
Operators/laborers
Production
Office/admin
Sales
Technicians
Professionals
Manager
s
1979–1989 1989–1999 1999–2007 2007–2012
14 Journal of Economic Perspectives
precisely, the figure plots 100 times log changes in employment, which are close
to equivalent to percentage points for small changes. Agricultural occupations
comprise no more than 2.2 percent of employment in this time interval, so this
omission has a negligible effect.) These ten occupations can be divided into
three groups. On the right-hand side of the figure are managerial, professional,
and technical occupations, which are highly educated and highly paid. Moving
leftward, the next four columns display employment growth in middle-skill occu-
pations, comprising sales; office and administrative support; production, craft
and repair; and operator, fabricator, and laborer. The leftmost three columns of
Figure 2 depict employment trends in service occupations, defined by the Census
Bureau as jobs that involve helping, caring for, or assisting others. The majority
of workers in service occupations have no post-secondary education, and average
hourly wages in service occupations are in most cases below the other seven occu-
pational categories.
As Figure 2 illustrates, the rapid employment growth in both high- and
low-education jobs has substantially reduced the share of employment accounted
for by “middle-skill” jobs. In 1979, the four middle-skill occupations (sales; office
and administrative workers; production workers; and operatives) accounted for
60 percent of employment. In 2007, this number was 49 percent, and in 2012, it
was 46 percent. The employment share of service occupations was essentially flat
between 1959 and 1979, and so their rapid growth since 1980 marks a sharp trend
reversal (Autor and Dorn 2013).
The polarization of employment across occupations is not unique to the
United States. Figure 3 plots changes in the share of employment between 1993
and 2010 within three broad sets of occupations—low-, middle-, and high-wage—
covering all nonagricultural employment in 16 European Union economies. In all
countries, middle-wage occupations declined as a share of employment while both
high-wage and low-wage occupations increased their shares of employment over
this 17-year period. While the US and EU data are not precisely comparable, the
US economy would fall roughly in the middle of the pack of this set of countries
in terms of its employment polarization. The comparability of these occupational
shifts across a large set of developed countries makes it likely that a common set
of forces contributes to these shared labor-market developments. Simultaneously,
the substantial differences among countries underscores that no single factor or
common cause explains the diversity of experiences across the United States and
the EuropeanUnion.
Does Employment Polarization Lead to Wage Polarization?
From the barbell shape of occupational employment growth depicted in
Figures 2 and 3, one might surmise that occupational polarization would also cata-
lyze wage polarization—that is, rising relative wages in both high-education, abstract
task-intensive jobs and in low-education, manual task-intensive jobs. However, this
Why Are There Still So Many Jobs? 15
reasoning does not take into account the role played by the three mitigating forces
discussed above: complementarity, demand elasticity, and labor supply.
Let’s first consider the effect of computerization on wages in abstract task-intensive
occupations such as managerial, professional, and technical occupations. These occu-
pations all draw upon large bodies of constantly evolving expertise: for example,
medical knowledge, legal precedents, sales data, financial analysis, programming
languages, and economic statistics. Information technology and computerization
should strongly complement workers performing abstract task-intensive jobs. By
dramatically lowering the cost and increasing the scope of information and analysis
available to them, computerization enables workers performing abstract tasks to
further specialize in their area of comparative advantage, with less time spent on
acquiring and crunching information, and more time spent on interpreting and
applying it. By the same token, information technology substitutes for many of the
Figure 3
Change in Occupational Employment Shares in Low, Middle, and High-Wage
Occupations in 16 EU Countries, 1993–2010
Source: Goos, Manning, and Salomons (2014, table 2).
Notes: High-paying occupations are corporate managers; physical, mathematical, and engineering
professionals; life science and health professionals; other professionals; managers of small enterprises;
physical, mathematical, and engineering associate professionals; other associate professionals; life science
and health associate professionals. Middle-paying occupations are stationary plant and related operators;
metal, machinery, and related trade work; drivers and mobile plant operators; office clerks; precision,
handicraft, craft printing, and related trade workers; extraction and building trades workers; customer
service clerks; machine operators and assemblers; and other craft and related trade workers. Low-paying
occupations are laborers in mining, construction, manufacturing, and transport; personal and protective
service workers; models, salespersons, and demonstrators; and sales and service elementary occupations.
−14.9%
−12.1%
−12.0%
−10.9%
−10.8%
−10.7%
−10.6%
−10.6%
−10.4%
−10.3%
−9.6%
−8.6%
−8.5%
−7.6%
−6.7%
−4.9%
−18%
−15%
−12%
−9%
−6%
−3%
0%
3%
6%
9%
12%
15%
Ireland
Belgium
Spain
United Kingdom
Luxembourg
Greece
Finland
Italy
Austria
Denmark
Sweden
France
Norway
Netherlands
Germany
Portugal
Low paying
Middle paying
High paying
16 Journal of Economic Perspectives
support occupations that these professions employ, including medical secretaries, para-
legals, and research assistants. Similarly, computerization and information technology
appears to allow “delayering” of management structures (Caroli and Van Reenen
2001). Arguably, many of the middle managers displaced by delayering performed
routine information-processing tasks.
If demand for the output of abstract task-intensive activities is inelastic, these
productivity gains might work to lower expenditure on these outputs, which could
mitigate wage gains. However, all outward evidence suggests that as technology has
boosted the output of the professions, demand for their services has more than kept
pace. Health care is an obvious example, but one can readily make similar argu-
ments about finance, law, engineering, research, and design.
What about reactions from labor supply? If workers could quickly move into
the highly educated professions, such a shift would mute earnings gains. But of
course, many professions require both college and graduate degrees, so the produc-
tion pipeline for new entrants is at least five to ten years in length. Indeed, young
US adults, particularly US males, have responded with remarkable sluggishness to
the rising educational premium over the last 30 years (Autor 2014). For example, in
1975, approximately 40 percent of hours worked by males with fewer than ten years
of experience (a group that has made the more recent choices about college) were
supplied by those with a college education. Forty years later in 2005, this share was
almost unchanged. For women workers with less than ten years of experience, the
share of total hours worked by those with a college education was 42 percent in 1982
but had risen to 53 percent by 2005. In the last decade, the share of hours worked by
those with less than ten years of experience and a college degree has increased for
both men and women: in 2012, it was 52 percent of hours for men in this group and
62 percent of the hours for women. Thus, while the stock of workers with college
and graduate degrees has certainly grown, the supply response has not been nearly
large enough to swamp the contemporaneous movements in labor demand.
Workers in abstract task-intensive occupations therefore benefit from informa-
tion technology via a virtuous combination of strong complementarities between
routine and abstract tasks, elastic demand for services provided by abstract
task-intensive occupations, and inelastic labor supply to these occupations over the
short and medium term. In combination, these forces mean that information tech-
nology should raise earnings in occupations that make intensive use of abstract tasks
and among workers who intensively supply them.
These same synergies do not apply to jobs that are intensive in manual tasks,
such as janitors and cleaners, vehicle drivers, security guards, flight attendants,
food service workers, and home health aides. Most manual task-intensive occupa-
tions are only minimally reliant on information or data processing for their core
tasks, and involve only limited opportunities for either direct complementarity
orsubstitution.
6
6
There are partial exceptions to this generalization: global positioning system satellites and scheduling
software allows truckers and delivery services to minimize wasted mileage; calendar, contact, and billing
David H. Autor 17
Aggregate evidence suggests that final demand for manual task-intensive
work—services in particular—is relatively price inelastic (Baumol 1967; Autor and
Dorn 2013). If so, productivity gains in manual task-intensive occupations that tend
to reduce their price per unit of service provided will not necessarily raise expendi-
ture on their outputs. On the other hand, demand for manual task-intensive work
appears to be relatively income elastic (Clark 1951; Mazzorali and Ragusa 2013),
so that rising aggregate incomes will tend to increase demand for these activities.
New technology and productivity growth in other areas may therefore indirectly raise
demand for manual task-intensive occupations by increasing societal income.
Labor supply to manual task-intensive occupations is intrinsically elastic, due
to their generally low education and training requirements. This insight does not
preclude the possibility that wages in manual tasks will rise, at least to some extent. As
Baumol (1967) observed, even absent productivity growth in technologically lagging
occupations, wages in these occupations must rise over time with societal income to
compensate workers for not entering other sectors (again, assuming that demand
for these activities is relatively inelastic). But it does suggest that wage increases in
these jobs will be restrained to some extent by the labor supply response, including
from workers displaced in other sectors of the economy.
Overall, manual task-intensive activities are at best weakly complemented by
computerization, do not benefit from elastic final demand, and face elastic labor
supply that tempers demand-induced wage increases. Thus, while information tech-
nology has strongly contributed to employment polarization measured in quantity of
jobs, we would not generally expect these employment changes to culminate in a
corresponding wage polarization except perhaps at certain times or in certain labor
markets. Indeed, in Autor and Dorn (2013), we present evidence that wages for
manual-task occupations rose during the 1990s when labor markets were extremely
tight, but after 2000, the expansion of manual task-intensive service occupations
accelerated while wages in these occupations fell.
For insight about the evolution of wage patterns, consider Figure 4. The hori-
zontal axis of this figure is based on a ranking of all 318 detailed occupations from
lowest to highest by their initial skill level, as measured by its 1979 mean hourly
occupational wage. These categories are weighted by their initial size, and then
grouped into 100 bins of equal size. The vertical axis of the figure then shows the
percentage change in wages over each of four periods across the skill distribution—
with the line smoothed for clarity. (Again, more precisely, the figure plots 100 times
log changes in employment, which is nearly equivalent to percentage points for
small changes.)
The right-hand two-thirds of Figure 4 look like the plots of employment polariza-
tion. From 1979 through 2007, wages rose consistently across the high-skill portion
software assists home health workers to manage data more effectively; and computerized ordering
systems enable food service workers to rapidly tally customer tabs. In a few years time, many retailers
may employ RFID “chip” technology that will scan purchases without needing a human checkout cashier
atall.
18 Journal of Economic Perspectives
of the figure, which is disproportionately made up of the abstract task-intensive
categories of professional, technical, and managerial occupations. By contrast, wage
growth in the middle-skill, typically routine task-intensive occupations was less rapid
and generally decelerated over time. For the low-education, manual task-intensive
occupations heavily represented on the left-hand side of Figure 4, in the 1980s,
wage growth was a little more rapid than in the middle-skill occupations—and in
the 1990s, it was much more rapid. However, that changed in the 2000s: while
Figure 2 showed that employment growth in these occupations exceeded that in
all other categories between 1999 and 2007, Figure 4 shows wage growth was gener-
ally negative in the low-skill percentiles, lower than in all other categories (Mishel,
Shierholz, and Schmitt 2013). During this time period, my strong hunch is that
the explanation is that declining employment in middle-skill routine task-intensive
Figure 4
Changes in Mean Wages by Occupational Skill Percentile among Full-Time,
Full-Year (FTFY) Workers, 1979–2012
(the y-axis plots 100 times log changes in employment, which is nearly equivalent to
percentage points for small changes)
Sources: Author, calculated using 1980, 1990, and 2000 Census IPUMS files; American Community Survey
combined file 2006–2008, American Community Survey 2012.
Notes: The figure plots changes in mean log wages over each period, by 1979 occupational skill percentile
rank using a locally weighted smoothing regression (bandwidth 0.8 with 100 observations), where skill
percentiles are measured as the employment-weighted percentile rank of an occupation’s mean log
wage in the Census IPUMS 1980 5 percent extract. The sample includes the working-age (1–64) civilian
non-institutionalized population with 48+ annual weeks worked and 35+ usual weekly hours. Weekly
wages are calculated as annual earnings divided by weeks worked.
−5
0
5
10
15
20
100 × Log Change in Mean FTFY Wage
0 20 40 60 80 100
Skill percentile (ranked by occupation’s 1979 mean log wage)
1979–1989 1989–1999 1999–2007 2007–2012
Why Are There Still So Many Jobs? 19
jobs led middle-skill workers—including new entrants, those displaced from routine
task-intensive jobs, and those who lost jobs during recession—to enter manual
task-intensive occupations instead (Smith 2013; Cortes, Jaimovich, Nekarda, and
Siu 2014; Foote and Ryan 2014).
A final set of facts illustrated by Figure 4 is that overall wage growth was anemic
throughout the 2000s, even prior to the Great Recession. Between 1999 and 2007,
real wage changes were negative below approximately the 15th percentile, and were
below 5 percentage points up to the 70th percentile of the distribution. Indeed,
wage growth was greater at all percentiles during both the 1980s and 1990s than in
the pre-recession 2000s.
7
Of course, wage growth was essentially zero at all percentiles
from 2007 to 2012.
Why are the rapidly rising earnings of the top 1 percent (as discussed in
Atkinson, Piketty, and Saez 2011, for example) not strongly evident in Figure 4?
One reason reflects substance; another is an artifact of the data. Substantively, the
plot depicts changes in earnings by occupational percentile rather than wage percen-
tile. Wage growth by occupational percentile is less concentrated than wage growth
across wage percentiles because the highest earners are found across a variety of
occupations. In addition, the very highest percentiles of earnings are censored in
public use Census and American Community Survey data files, which further masks
earnings gains at extreme quantiles.
The Recent Slowdown in the Growth of High-Skill Occupations
The hypothesis that automation and information technology has led to occu-
pational and, to a lesser degree, wage polarization in the US labor force can explain
some key features of the US and the cross-national data. But reality invariably proves
more complicated than any single theory anticipates.
For my thesis linking technological change to occupational change, one
concern is the unexplained deceleration of employment growth in abstract
task-intensive occupations after 2000 (Beaudry, Green, and Sand 2014, forth-
coming; Mishel, Shierholz, and Schmitt 2013). Figure 5 follows the format of
Figure 4 but instead of showing (approximate) percentage changes in wages on
the vertical axis, it shows percentage changes in the employment share of the jobs
ranked by their skill level in 1979. Since the sum of shares must equal one at any
time period, the changes in these shares across the decades must total zero, and
thus, the height at each skill percentile measures the growth in each occupation’s
employment relative to thewhole.
7
Because the 2000–2007 interval is two years shorter than the 1979–1989 period, one should multiply
the later changes by 1.25 to put them on the same temporal footing. But even after making such an
adjustment, wage growth was still considerably weaker at all percentiles from 2000–2007 than in the
earlier two decades.
20 Journal of Economic Perspectives
Figure 5 contributes three nuances to the occupational polarization story
above. First, the pace of employment gains in low-wage, manual task-intensive jobs
has risen successively across periods, as shown at the left-hand side of the figure.
Second, the occupations that are losing employment share appear to be increas-
ingly drawn from higher ranks of the occupational distribution. For example, the
highest ranked occupation to lose employment share during the 1980s lay at approx-
imately the 45th percentile of the skill distribution. In the final two subperiods, this
rank rose still further to above the 75th percentile—suggesting that the locus of
displaced middle-skill employment is moving into higher-skilled territories. Third,
growth of high-skill, high-wage occupations (those associated with abstract work)
decelerated markedly in the 2000s, with no relative growth in the top two deciles
of the occupational skill distribution during 1999 through 2007, and only a modest
recovery between 2007 and 2012. Stated plainly, the growth of occupational employ-
ment across skill levels looks U-shaped earlier in the period, with gains at low-skill
and high-skill levels. By the 2000s, the pattern of occupational employment across
Figure 5
Smoothed Employment Changes by Occupational Skill Percentile, 1979–2012
Sources: Author, calculated using 1980, 1990, and 2000 Census Integrated Public Use Microdata Series
(IPUMS) files; American Community Survey combined file 2006–2008, American Community Survey 2012.
Notes: The figure plots changes in employment shares by 1980 occupational skill percentile rank using a
locally weighted smoothing regression (bandwidth 0.8 with 100 observations), where skill percentiles are
measured as the employment-weighted percentile rank of an occupation’s mean log wage in the Census
IPUMS 1980 5 percent extract. Employment in each occupation is calculated using workers’ hours of
annual labor supply times the Census sampling weights. Consistent occupation codes for Census years
1980, 1990, and 2000, and 2008 are from Autor and Dorn (2013).
−.1
0
.1
.2
100 × Change in Employment Share
0 20 40 60 80 100
Skill percentile (ranked by occupation’s 1979 mean log wage)
1979–1989 1989–1999 1999–2007 2007–2012
David H. Autor 21
skill levels began to resemble a downward ramp. In Autor (2015), I present a more
detailed breakdown of these patterns, and in particular suggest that the set of
abstract task-intensive jobs is not growing as rapidly as the potential supply of highly
educated workers.
What explains the slowing growth of abstract task-intensive employment?
One interpretation is that automation, information technology, and technological
progress in general are encroaching upward in the task domain and beginning to
substitute strongly for the work done by professional, technical, and managerial
occupations. While one should not dismiss this possibility out of hand, it doesn’t
fit well with the pattern of computer and software investment. If information tech-
nology is increasingly replacing workers high in the skill distribution, one would
expect a surge of corporate investment in computer hardware and software. Instead,
Figure 6 shows that in early 2014, information processing equipment and software
investment was only 3.5 percent of GDP, a level last seen in 1995 at the outset of
the “dot-com” era. To me, the evidence in Figure 6 suggests a temporary disloca-
tion of demand for information technology capital during the latter half of the
1990s, followed by a sharp correction after 2000. I suspect that the huge falloff in
Figure 6
Private Fixed Investment in Information Processing Equipment and Software as a
Percentage of Gross Domestic Product, 1949–2014
Source: FRED, Federal Bank of St. Louis. http://research.stlouisfed.org/fred2/graph/?g=GXc (accessed
8/3/2014).
0.0%
1.0%
1.5%
2.0%
2.5%
3.0%
3.5%
4.0%
4.5%
5.0%
1949 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 2004 2009 2014
0.5%
22 Journal of Economic Perspectives
information investment may have dampened innovative activity and demand for
high-skilled workers more broadly.
As noted earlier, technological change is far from the only factor affecting US
labor markets in the last 15 years. For example, the deceleration of wage growth
and changes in occupational patterns in the US labor market after 2000, and
further after 2007, is surely associated to some extent with two types of macro-
economic events. First, there are the business cycle effects—the bursting of the
“dot-com” bubble in 2000, and the collapse of the housing market and the ensuing
financial crisis in 2007–2008—both of which curtailed investment and innovative
activity. Second, there are the employment dislocations in the US labor market
brought about by rapid globalization, particularly the sharp rise of import pene-
tration from China following its accession to the World Trade Organization in
2001 (Autor, Dorn, and Hanson 2013; Pierce and Schott 2012; Acemoglu, Autor,
Dorn, Hanson, and Price forthcoming). China’s rapid rise to a premier manufac-
turing exporter had far-reaching impacts on US workers, reducing employment
in directly import-competing US manufacturing industries and depressing labor
demand in both manufacturing and nonmanufacturing sectors that served as
upstream suppliers to these industries.
Of course, these forces are in various ways linked with the spread of automa-
tion and technology. Advances in information and communications technologies
have changed job demands in US workplaces directly and also indirectly, by making
it increasingly feasible and cost-effective for firms to source, monitor, and coordi-
nate complex production processes at disparate locations worldwide and altering
competitive conditions for US manufacturers and workers. This multidimensional
complementarity among causal factors makes it both conceptually and empirically
difficult to isolate the “pure” effect of any one factor.
Polanyi’s Paradox: Will It Be Overcome?
Automation, complemented in recent decades by the exponentially increasing
power of information technology, has driven changes in productivity that have
disrupted labor markets. This essay has emphasized that jobs are made up of many
tasks and that while automation and computerization can substitute for some of
them, understanding the interaction between technology and employment requires
thinking about more than just substitution. It requires thinking about the range
of tasks involved in jobs, and how human labor can often complement new tech-
nology. It also requires thinking about price and income elasticities for different
kinds of output, and about labor supply responses.
The tasks that have proved most vexing to automate are those demanding flexi-
bility, judgment, and common sense—skills that we understand only tacitly. Ireferred
to this constraint above as Polanyi’s paradox. In the past decade, computerization
and robotics have progressed into spheres of human activity that were considered
off limits only a few years earlier—driving vehicles, parsing legal documents, even
Why Are There Still So Many Jobs? 23
performing agricultural field labor. Is Polanyi’s paradox soon to be at least mostly
overcome, in the sense that the vast majority of tasks will soon be automated?
8
My reading of the evidence suggests otherwise. Indeed, Polanyi’s paradox helps
to explain what has not yet been accomplished, and further illuminates the paths by
which more will ultimately be accomplished. Specifically, I see two distinct paths that
engineering and computer science can seek to traverse to automate tasks for which we
“do not know the rules”: environmental control and machine learning. The first path
circumvents Polanyi’s paradox by regularizing the environment, so that comparatively
inflexible machines can function semi-autonomously. The second approach inverts
Polanyi’s paradox: rather than teach machines rules that we do not understand, engi-
neers develop machines that attempt to infer tacit rules from context, abundant data,
and applied statistics.
Environmental Control
Most automated systems lack flexibility—they are brittle. Modern automobile
plants, for example, employ industrial robots to install windshields on new vehicles
as they move through the assembly line. But aftermarket windshield replacement
companies employ technicians, not robots, to install replacement windshields.
Evidently, the tasks of removing a broken windshield, preparing the windshield frame
to accept a replacement, and fitting a replacement into that frame demand more
real-time adaptability than any contemporary robot can cost-effectivelyapproach.
The distinction between assembly line production and the in-situ repair
highlights the role of environmental control in enabling automation. Engineers
can in some cases radically simplify the environment in which machines work to
enable autonomous operation, as in the familiar example of a factory assembly
line. Numerous examples of this approach to environmental regularization are
so ingrained in daily technology that they escape notice, however. To enable the
operation of present-day automobiles, for example, humanity has adapted the natu-
rally occurring environment by leveling, re-grading, and covering with asphalt a
nontrivial percentage of the earth’s land surface.
9
The ongoing automation of warehouses provides another example. Large
online retailers, such as Amazon.com, Zappos.com, and Staples, operate systems of
warehouses that have traditionally employed legions of dexterous, athletic “pickers,”
who run and climb through shelves of typically non-air-conditioned warehouses to
locate, collect, box, label, and ship goods. There is at present no cost-effective robotic
8
For a glimpse of the view that just about anything can now be computerized, see the widely cited
(albeit unpublished) article by the economists Carl Frey and Michael Osborne, who write (2013, p. 24)
that, “recent developments in ML [machine learning] and MR [mobile robotics], building upon big
data, allow for pattern recognition, and thus enable computer capital to rapidly substitute for labour
across a wide range of non-routine tasks. Yet some inhibiting engineering bottlenecks to computerization
persist. Beyond these bottlenecks, however, we argue that it is largely already technologically possible to
automate almost any task, provided that sufficient amounts of data are gathered for pattern recognition.”
9
According to Wikipedia, so-called impervious surfaces (mostly roads and parking lots) cover 43,000
square miles of land in the lower 48 United States—roughly equal to the land area of the state of Ohio
(http://en.wikipedia.org/wiki/Impervious_surface, accessed 8/4/2014).
24 Journal of Economic Perspectives
facsimile for these human pickers. The job’s steep requirements for flexibility, object
recognition, physical dexterity, and fine motor coordination are tooformidable.
But large components of warehousing can be automated, as demonstrated
by Kiva Systems, a robotic warehousing startup that was purchased by Amazon in
2012. The core of the Kiva system is a dispatch program that oversees the flow of all
goods through the warehouse, coordinating the work of robots, which carry shelves,
with the work of humans. As objects arrive at the facility for stocking, the dispatch
software directs robots to transport and line up empty shelves to a loading area,
where human stockers place merchandise on shelves. Robots then carry the loaded
shelves back to a storage warehouse, where the dispatch software directs their place-
ment to optimize product availability for expected product demand. As new orders
arrive, the dispatch software sends robots to retrieve shelves and lines them up in
a packing area. Then a human picker, directed by a laser pointer controlled by the
dispatch software, takes objects from the assembled shelves, packs them in shipping
boxes, applies a shipping label, and drops the package in a chute for delivery. As
items are picked, the robots take the shelves away until needed again for packing
or restocking. Thus, in a Kiva-operated warehouse, robots handle only the routine
task of moving shelves across a level surface; workers handle merchandise; and the
dispatch software coordinates the activity.
While Kiva Systems provides a particularly clear example of exploiting envi-
ronmental control to extend the reach of automation, the same principle is often
lurking behind more sophisticated packaging. Perhaps the least recognized—and
most mythologized—is the self-driving Google Car. Computer scientists sometimes
remark that the Google car does not drive on roads, but rather on maps. AGoogle
car navigates through the road network primarily by comparing its real-time
audio-visual sensor data against painstakingly hand-curated maps that specify the
exact locations of all roads, signals, signage, and obstacles. The Google car adapts
in real time to obstacles, such as cars, pedestrians, and road hazards, by braking,
turning, and stopping. But if the car’s software determines that the environment in
which it is operating differs from the environment that has been preprocessed by its
human engineers—when it encounters an unexpected detour or a crossing guard
instead of a traffic signal—the car requires its human operator to take control. Thus,
while the Google car appears outwardly to be adaptive and flexible, it is somewhat
akin to a train running on invisible tracks.
These examples highlight both the limitations of current technology to
accomplish nonroutine tasks, and the capacity of human ingenuity to surmount
some of these obstacles by re-engineering the environment in which work tasks
areperformed.
Machine Learning
Polanyi’s paradox—“we know more than we can tell”—presents a challenge for
computerization because, if people understand how to perform a task only tacitly
and cannot “tell” a computer how to perform the task, then seemingly programmers
cannot automate the task—or so the thinking has gone. But this understanding
David H. Autor 25
is shifting rapidly due to advances in machine learning. Machine learning applies
statistics and inductive reasoning to supply best-guess answers where formal proce-
dural rules are unknown. Where engineers are unable to program a machine to
“simulate” a nonroutine task by following a scripted procedure, they may neverthe-
less be able to program a machine to master the task autonomously by studying
successful examples of the task being carried out by others. Through a process of
exposure, training, and reinforcement, machine learning algorithms may poten-
tially infer how to accomplish tasks that have proved dauntingly challenging to
codify with explicit procedures.
As a concrete example, consider the task of visually identifying a chair (discussed
in Autor, forthcoming). An engineer applying a conventional rules-based program-
ming paradigm might attempt to specify what features of an object qualify an object
as a chair—it possesses legs, arms, a seat, and a back, for example. But one would
soon discover that many chairs do not possess all of these features (for example,
some chairs have no back, or no arms). If the engineer then relaxed the required
feature set accordingly (chair back optional), the included set would grow to encom-
pass many objects that are not chairs, such as small tables. The canonical approach
to recognizing objects by pre-specifying requisite features—and more sophisticated
variants of this approach—would likely have very high misclassification rates. Yet,
any grade-school child could perform this task with high accuracy. What does the
child know that the rules-based procedure does not? Unfortunately, we cannot
enunciate precisely what the child knows—and this is precisely Polanyi’s paradox.
Machine learning potentially circumvents this problem. Relying on large data-
bases of so-called “ground truth”—a vast set of curated examples of labeled objects—a
machine learning algorithm attempts to infer what attributes of an object make it
more or less likely to be designated a chair. This process is called “training.” When
training is complete, the machine can apply this statistical model to attempt to iden-
tify chairs that are distinct from those in the original dataset. If the statistical model is
sufficiently good, it may be able to recognize chairs that are somewhat distinct from
those in the original training data, like chairs of different shapes, materials, or dimen-
sions. Machine learning does not require an explicit physical model of “chairness.” At
its core, machine learning is an atheoretical brute force technique—what psycholo-
gists call “dustbowl empiricism”—requiring only large training databases, substantial
processing power, and, of course, sophisticated software.
10
How well does machine learning work in practice? If you use a search engine
or Google Translate, operate a smartphone with voice commands, or follow movie
suggestions from Netflix, you can assess for yourself how successfully these tech-
nologies function. For example, if the majority of users who recently searched for
the terms “degrees bacon” clicked on links for Kevin Bacon rather than links
for best bacon cooking temperatures, the search engine would tend to place the
Kevin Bacon links higher in the list of results. My general observation is that
10
Varian (2014) provides an introduction to machine learning techniques for economists.
26 Journal of Economic Perspectives
the tools are inconsistent: uncannily accurate at times; typically only so-so; and
occasionally unfathomable. Moreover, an irony of machine learning algorithms is
that they also cannot “tell” programmers why they do what they do. IBM’s Watson
computer famously triumphed in the trivia game of Jeopardy against champion human
opponents. Yet Watson also produced a spectacularly incorrect answer during its
winning match. Under the category of US Cities, the question was, “Its largest airport
was named for a World WarII hero; its second largest, for a World WarII battle.”
Watson’s proposed answer was Toronto, a city in Canada. Even leading-edge accom-
plishments in this domain can appear somewhat underwhelming. A 2012 New York
Times article (Markoff 2012) described Google’s X Lab’s recent project (Le et al.
2012) to apply a neural network of 16,000 processors to identify images of cats on
YouTube. The article’s headline ruefully poses the question, “How Many Computers
to Identify a Cat? 16,000.”
Since the underlying technologies—the software, hardware, and training data—
are all improving rapidly (Andreopouos and Tsotsos 2013), one should view these
examples as prototypes rather than as mature products. Some researchers expect
that as computing power rises and training databases grow, the brute force machine
learning approach will approach or exceed human capabilities. Others suspect that
machine learning will only ever “get it right” on average, while missing many of
the most important and informative exceptions. Ultimately, what makes an object a
chair is that it is purpose-built for a human being to sit upon. Machine-learning algo-
rithms may have fundamental problems with reasoning about “purposiveness” and
intended uses, even given an arbitrarily large training database of images (Grabner,
Gall, and Van Gool 2011). One is reminded of Carl Sagan’s (1980, p. 218) remark,
“If you wish to make an apple pie from scratch, you must first invent the universe.”
Conclusions
Major newspaper stories offer fresh examples daily of technologies that substi-
tute for human labor in an expanding—although still circumscribed—set of tasks.
The offsetting effects of complementarities and rising demand in other areas are,
however, far harder to identify as they occur. My own prediction is that employ-
ment polarization will not continue indefinitely (as argued in Autor 2013). While
some of the tasks in many current middle-skill jobs are susceptible to automation,
many middle-skill jobs will continue to demand a mixture of tasks from across the
skill spectrum. For example, medical support occupations—radiology techni-
cians, phlebotomists, nurse technicians, and others—are a significant and rapidly
growing category of relatively well-remunerated, middle-skill employment. Most of
these occupations require mastery of “middle-skill” mathematics, life sciences, and
analytical reasoning. They typically require at least two years of postsecondary voca-
tional training, and in some cases a four-year college degree or more. This broad
description also fits numerous skilled trade and repair occupations, including
plumbers, builders, electricians, heating/ventilating/air-conditioning installers, and
Why Are There Still So Many Jobs? 27
automotive technicians. It also fits a number of modern clerical occupations that
provide coordination and decision-making functions, rather than simply typing and
filing, like a number of jobs in marketing. There are also cases where technology is
enabling workers with less esoteric technical mastery to perform additional tasks: for
example, the nurse practitioner occupation that increasingly performs diagnosing
and prescribing tasks in lieu of physicians.
I expect that a significant stratum of middle-skill jobs combining specific voca-
tional skills with foundational middle-skills levels of literacy, numeracy, adaptability,
problem solving, and common sense will persist in coming decades. My conjec-
ture is that many of the tasks currently bundled into these jobs cannot readily
be unbundled—with machines performing the middle-skill tasks and workers
performing only a low-skill residual—without a substantial drop in quality. This
argument suggests that many of the middle-skill jobs that persist in the future will
combine routine technical tasks with the set of nonroutine tasks in which workers
hold comparative advantage: interpersonal interaction, flexibility, adaptability,
and problem solving. In general, these same demands for interaction frequently
privilege face-to-face interactions over remote performance, meaning that these
same middle-skill occupations may have relatively low susceptibility to offshoring.
Lawrence Katz memorably titles workers who virtuously combine technical and
interpersonal tasks as “the new artisans” (see Friedman 2010), and Holzer (2015)
documents that “new middle skill jobs” are in fact growing rapidly, even as tradi-
tional production and clerical occupations contract.
11
This prediction has one obvious catch: the ability of the US education and
job training system (both public and private) to produce the kinds of workers
who will thrive in these middle-skill jobs of the future can be called into question.
In this and other ways, the issue is not that middle-class workers are doomed by
automation and technology, but instead that human capital investment must be
at the heart of any long-term strategy for producing skills that are complemented
by rather than substituted for by technological change. In 1900, the typical young,
native-born American had only a common school education, about the equivalent
of sixth to eighth grades. By the late 19th century, many Americans recognized that
this level of schooling was inadequate: farm employment was declining, industry was
rising, and their children would need additional skills to earn a living. The United
States responded to this challenge over the first four decades of the 20th century by
becoming the first nation in the world to deliver universal high school education to
its citizens (Goldin and Katz 2008). Tellingly, the high school movement was led by
the farm states. Societal adjustments to earlier waves of technological advancement
were neither rapid, automatic, nor cheap. But they did pay off handsomely.
11
A creative paper by Lin (2011) studies the growth of “new work” by documenting the differential
growth of US employment in newly introduced Census occupation codes during the 1980s and 1990s
in high-education and high-technology cities. New occupational titles are generally clustered across two
categories: those associated with using new technologies such as web developer or database adminis-
trator; and novel personal services, such as personal chefs and stylists.
28 Journal of Economic Perspectives
A final point, typically neglected in recent dismal prophesies of machine-human
substitution, is that if human labor is indeed rendered superfluous by automation,
then our chief economic problem will be one of distribution, not of scarcity. The
primary system of income distribution in market economies is rooted in labor scar-
city; citizens possess (or acquire) a bundle of valuable “human capital” that, due to
its scarcity, generates a flow of income over the career path. If machines were in fact
to make human labor superfluous, we would have vast aggregate wealth but a serious
challenge in determining who owns it and how to share it. One might presume that
with so much wealth at hand, distribution would be relatively straightforward to
resolve. But history suggests that this prediction never holds true. There is always
perceived scarcity and ongoing conflict over distribution, and I do not expect that
this problem will become any less severe as automation advances. Are we actually
on the verge of throwing off the yoke of scarcity so that our primary economic
challenge soon becomes one of distribution? Here, I recall the observations of econ-
omist, computer scientist, and Nobel laureate Herbert Simon (1966), who wrote
at the time of the automation anxiety of the 1960s: “Insofar as they are economic
problems at all, the world’s problems in this generation and the next are problems
of scarcity, not of intolerable abundance. The bogeyman of automation consumes
worrying capacity that should be saved for real problems . . .” A half century on,
Ibelieve the evidence favors Simon’s view.
■ This paper draws from an essay prepared for the Federal Reserve Bank of Kansas City’s
economic policy symposium on “Re-Evaluating Labor Market Dynamics,” August 21–23,
2014, in Jackson Hole, Wyoming (Autor 2015) as well as the essay “The Paradox of
Abundance: Automation Anxiety Returns” (Autor forthcoming). I thank Erik Brynjolfsson,
Chris Foote, Frank Levy, Lisa Lynch, Andrew McAfee, Brendan Price, Seth Teller, Dave
Wessel, participants in the MIT CSAIL/Economists Lunch Seminar, and the editors of this
journal for insights that helped to shape my thinking on this subject. I thank Sookyo Jeong and
Brendan Price for superb research assistance.
References
Acemoglu, Daron, David H. Autor, David Dorn,
Gordon Hanson, and Brendan Price. Forthcoming.
“Import Competition and the Great U.S. Employ-
ment Sag of the 2000s.” Journal of Labor Economics.
Akst, Daniel. 2013. “What Can We Learn from
Past Anxiety over Automation?” The Wilson Quar-
terly, Summer.
Alpert, Andrew, and Jill Auyer. 2003. “Evaluating
the BLS 1988–2000 Employment Projections.”
Monthly Labor Review, October, pp. 13–37.
Andreopoulos, Alexander, and John K. Tsotsos.
2013. “50 Years of Object Recognition: Directions
Forward.” Computer Vision and Image Understanding
117(8): 827–91.
David H. Autor 29
Atkinson, Anthony B., Thomas Piketty, and
Emmanuel Saez. 2011. “Top Incomes in the Long
Run of History.” Journal of Economic Literature
49(1): 3–71.
Autor, David H. 2013. “The ‘Task Approach’ to
Labor Markets: An Overview.” Journal for Labour
Market Research 46(3): 185–99.
Autor, David H. 2014. “Skills, Education, and
the Rise of Earnings Inequality among the ‘Other
99 Percent.’” Science 344(6186): 843–51.
Autor, David H. 2015. “Polanyi’s Paradox and
the Shape of Employment Growth.” In Re-Evalu-
ating Labor Market Dynamics, pp. 129–79. Federal
Reserve Bank of Kansas City.
Autor, David H. Forthcoming. “The Paradox
of Abundance: Automation Anxiety Returns.” In
Performance and Progress: Essays on Capitalism, Busi-
ness and Society, edited by Subramanian Rangan.
London: Oxford University Press.
Autor, David H., and David Dorn. 2013. “The
Growth of Low-Skill Service Jobs and the Polariza-
tion of the US Labor Market.” American Economic
Review 103(5): 1553–97.
Autor, David H., David Dorn, and Gordon H.
Hanson. 2013. “The China Syndrome: Local Labor
Market Effects of Import Competition in the
United States.” American Economic Review 103(6):
2121–68.
Autor, David H., David Dorn, and Gordon H.
Hanson. 2015. “Untangling Trade and Technology:
Evidence from Local Labor Markets.” Economic
Journal 125(584): 621–46.
Autor, David H., Lawrence F. Katz, and Melissa
S. Kearney. 2006. “The Polarization of the U.S.
Labor Market.” American Economic Review 96(2):
189–94.
Autor, David H., Lawrence F. Katz, and Melissa
S. Kearney. 2008. “Trends in U.S. Wage Inequality:
Revising the Revisionists.” Review of Economics and
Statistics 90(2): 300–323.
Autor, David H., Frank Levy, and Richard J.
Murnane. 2003. “The Skill Content of Recent
Technological Change: An Empirical Exploration.”
Quarterly Journal of Economics 118(4): 1279–1333.
Baumol, William J. 1967. “Macroeconomics
of Unbalanced Growth: The Anatomy of Urban
Crisis.” American Economic Review 57(3): 415–26.
Beaudry, Paul, David A. Green, and Benjamin
M. Sand. 2014. “The Declining Fortunes of the
Young since 2000.” American Economic Review
104(5): 381–86.
Beaudry, Paul, David A. Green, and Benjamin
M. Sand. Forthcoming. “The Great Reversal in the
Demand for Skill and Cognitive Tasks.” Journal of
Labor Economics.
Bessen, James. 2015. “Toil and Technology.”
Finance and Development 52(1).
Bowen, Harold R. (Chairman). 1966. “Report
of the National Commission on Technology,
Automation, and Economic Progress: Volume I.”
Washington: U.S. Government Printing Office.
Brynjolfsson, Erik, and Andrew McAfee. 2014.
The Second Machine Age: Work, Progress, and Pros-
perity in a Time of Brilliant Technologies. New York
and London: W.W. Norton & Company.
Caroli, Eve, and John Van Reenen. 2001. “Skill-
Biased Organizational Change? Evidence from
a Panel of British and French Establishments.”
Quarterly Journal of Economics 116(4): 1449–92.
Clark, Colin. 1951. The Conditions of Economic
Progress, 2nd edition. London: Macmillan.
Cortes, Guido Matias, Nir Jaimovich,
Christopher J. Nekarda, and Henry E. Siu.
2014. “The Micro and Macro of Disappearing
Routine Jobs: A Flows Approach.” NBER Working
Paper20307, July.
Douglas, Paul H. 1930. Real Wages in the United
States, 1890–1926. New York: Houghton Mifflin
Company.
Foote, Christopher L., and Richard W. Ryan.
2014. “Labor-Market Polarization over the Busi-
ness Cycle.” Public Policy Discussion Paper 12-8,
Federal Reserve Bank of Boston, April.
Frey, Carl Benedikt, and Michael A. Osborne.
2013. “The Future of Employment: How Suscep-
tible are Jobs to Computerization?” Oxford Martin
School, September.
Friedman, Thomas L. 2010. “The Election That
Wasn’t.” New York Times, October 23.
Goldin, Claudia, and Lawrence F. Katz.
2008. The Race between Education and Technology.
Cambridge: Harvard University Press.
Goos, Maarten, and Alan Manning. 2003. “Lousy
and Lovely Jobs: The Rising Polarization of Work
in Britain.” Center for Economic Performance
Discussion Papers DP0604, December.
Goos, Maarten, and Alan Manning. 2007.
“Lousy and Lovely Jobs: The Rising Polarization of
Work in Britain.” Review of Economics and Statistics
89(1): 118–33.
Goos, Maarten, Alan Manning, and Anna
Salomons. 2014. “Explaining Job Polarization:
Routine-Biased Technological Change and Offshor-
ing.” American Economic Review 104(8): 2509–26.
Grabner, Helmut, Juergen Gall, and Luc Van
Gool. 2011. “What Makes a Chair a Chair?” In
Computer Vision and Pattern Recognition 2011, IEEE
Conference, June 20–25, pp. 1529–36.
Graetz, Georg, and Guy Michaels. 2015. “Robots
at Work.” IZA Discussion Papers 8938, Institute for
the Study of Labor (IZA).
Herald Press, The. 1966. “Skirting the Automa-
tion Question.” February 7, p. 2. (The Herald Press
of St. Joseph, Michigan.)
30 Journal of Economic Perspectives
Hoffman, William, and Leo Furcht. 2014. The
Biologist’s Imagination: Innovation in the Biosciences.
New York, NY: Oxford University Press.
Holzer, Harry J. 2015. “Job Market Polarization
and U.S. Worker Skills: A Tale of Two Middles.”
Brookings Institution Economic Studies Working
Paper, April 6.
Hsieh, Chang-Tai, and Enrico Moretti. 2003.
“Can Free Entry Be Inefficient? Fixed Commis-
sions and Social Waste in the Real Estate Industry.”
Journal of Political Economy 111(5): 1076–1122.
Jackson, Kristin. 1993. “The World’s First Motel
Rests upon Its Memories.” Seattle Times, April 25.
Jevons, William Stanley. 1865. The Coal Question;
An Inquiry Concerning the Progress of the Nation, and
the Probable Exhaustion of Our Coal Mines. London:
Macmillan and Co.
Katz, Lawrence F., and Robert A. Margo. 2014.
“Technical Change and the Relative Demand for
Skilled Labor: The United States in Historical
Perspective.” In Human Capital in History, edited by
Leah Platt Bouston, Carola Frydman, and Robert
A. Margo, 15–57. University of Chicago Press.
Kremer, Michael. 1993. “The O-Ring Theory
of Economic Development.” Quarterly Journal of
Economics 108(3): 551–75.
Le, Quoc V., Marc’ Aurelio Ranzato, Rajat
Monga, Matthieu Devin, Kai Chen, Greg S.
Corrado, Jeff Dean, and Andrew Y. Ng. 2012.
“Building High-Level Features Using Large Scale
Unsupervised Learning.” In Proceedings of the
29th International Conference on Machine Learning,
June 26–July 1, Edinburgh, Scotland, UK.
Levy, Frank and Richard J. Murnane. 2004. The
New Division of Labor: How Computers Are Creating
the Next Job Market. Princeton University Press.
Lin, Jeffrey. 2011. “Technological Adaptation,
Cities, and New Work.” Review of Economics and
Statistics 93(2): 554–74.
Markoff, John. 2012. “How Many Computers
to Identify a Cat? 16,000.” New York Times, June 25.
Mazzolari, Francesca, and Giuseppe Ragusa.
2013. “Spillovers from High-Skill Consumption to
Low-Skill Labor Markets.” Review of Economics and
Statistics 95(1): 74–86.
Michaels, Guy, Ashwini Natraj, and John Van
Reenen. 2014. “Has ICT Polarized Skill Demand?
Evidence from Eleven Countries over Twenty-Five
Years.” Review of Economics and Statistics 96(1): 60–77.
Mishel, Lawrence, Heidi Shierholz, and John
Schmitt. 2013. “Don’t Blame the Robots: Assessing
the Job Polarization Explanation of Growing
Wage Inequality.” EPI-CEPR Working Paper,
November 19.
Moravec, Hans.1988.Mind Children: The Future
of Robot and Human Intelligence. Harvard University
Press.
Nordhaus, William D. 2007. “Two Centuries
of Productivity Growth in Computing.” Journal of
Economic History 67(1): 17–22.
Pierce, Justin R., and Peter K. Schott. 2012.
“The Surprisingly Swift Decline of U.S. Manufac-
turing Employment.” NBER Working Paper 18655.
Polanyi, Michael. 1966. The Tacit Dimension.
New York: Doubleday.
Ruggles, Steven, Trent J. Alexander, Katie
Genadek, Ronald Goeken, Matthew B. Schroeder,
and Matthew Sobek. 2010. Integrated Public Use
Microdata Series: Version 5.0 [Machine-readable
database]. University of Minnesota.
Sachs, Jeffrey D., Seth G. Benzell, Guillermo
LaGarda. 2015. “Robots: Curse or Blessing? A Basic
Framework.” NBER Working Paper 21091, April.
Sachs, Jeffrey D., and Laurence J. Kotlikoff.
2012. “Smart Machines and Long-Term Misery.”
NBER Working Paper 18629, December.
Sagan, Carl. 1980. Cosmos. New York: Random
House.
Simon, Herbert. A. 1966. “Automation” (a letter
in response to “Where Do We Go From Here?”
March 17, 1966 issue) New York Review of Books,
May 26.
Smith, Christopher L. 2013. “The Dynamics
of Labor Market Polarization.” Federal Reserve
Board, Finance and Economics Discussion Series
No. 2013-57, August.
TIME. 1961. “The Automation Jobless.”
February 24.
US Bureau of the Census. 1949. Historical Statis-
tics of the United States, 1789–1945. U.S. Government
Printing Office, Series D 134-144.
Varian, Hal R. 2014. “Big Data: New Tricks for
Econometrics.” Journal of Economic Perspectives
28(2): 3–28.
