by Ajay Agrawal
themselves in the past in successive waves of added productivity growth, a
pattern that could repeat itself in the future (Syverson 2013).
12.2 Past Innovations Have Sometimes Increased Inequality—
and the Indications Suggest AI Could Be More of the Same
Advanced economies have seen vast amounts of innovation in the last
three centuries. Most of the kinds of jobs that existed in the 1700s do not
exist today, but jobs no one could have imagined then have taken their place.
As a result, over long periods of time it has generally been the case that about
95 percent of the people in the United States who want a job at a given point
in time can fi nd one—despite massive changes in technology.
Although labor markets do not function like the stylized models for a
commodity like wheat that populate economics textbooks, within broad
parameters the basic operation of supply and demand is the mechanism that
makes sure that just about everyone who wants a job can fi nd one. For this
to happen, however, wages need to adjust to make supply equal to demand.
In recent decades, much of that adjustment in wages has been in the form of
a large decline in wages for low- skill workers relative to high- skill workers.
If Automation in the Future Looks Like Automation in the Past 321
Fig. 12.3 Share of jobs with highly automatable skills by education
Source: Arntz, Gregory, and Zierahn (2016) calculations based on the Survey of Adult Skills (PIAAC 2012).
From 1975 until 2016, those with a high school degree watched their relative
wages fall from over 70 percent of the amount earned by full- time, full- year
workers with at least a college degree to just over 50 percent.
The worry is not that this time could be diff erent when it comes to AI,
but that this time could be the same as what we have experienced over the
past several decades. The traditional argument that we do not need to worry
about the robots taking our jobs still leaves us with the worry that the only
reason we will still have our jobs is because we are willing to do them for
lower wages.
The share of jobs that are threatened by future automation is fi ercely
debated, with estimates ranging from 9 percent by the Organisation for
Economic Co- operation and Development ([OECD]; Arntz, Gregory, and
Zierahn 2016, to 50 percent by Carl Frey and Michael Osborne 2013). While
this question is important, there is less ambiguity on the wages/ skills gradi-
ent of the jobs or tasks that are most likely to be substituted for by automa-
tion. The OECD researchers, for example, found that 44 percent of jobs with
less than a high school degree had highly automatable skills, as compared to
only 1 percent of jobs with a college degree, as shown in fi gure 12.3.
This is very similar to the gradient found in Frey and Osborne’s work.
The Council of Economic Advisers (Executive Offi
ce of the President 2016)
sorted the Frey and Osborne occupations at risk of automation by wages
and found that it ranged from 83 percent of occupations making less than
$20 an hour to only 4 percent of occupations making more than $40 per
hour, as shown in fi gure 12.4.
Since wages and skills are correlated, this means a large decline in the
322 Jason Furman
demand for lower- skill jobs and little decline in the demand for higher- skill
jobs. This result points to a shift in the impact of automation on the labor
market. At points in the past, automation led to a so-called polarization of
the labor market because jobs requiring a moderate skill level—which his-
torically included bookkeepers, clerks, and certain assembly- line workers—
were easier to routinize, although more recently that process of polariza-
tion appears to have stopped (Autor 2014; Schmitt, Schierholz, and Mishel
2013). Conversely, higher- skill jobs that use problem- solving capabilities,
intuition, and creativity, as well as lower- skill jobs that require situational
adaptability and in-person interactions, were less easy to routinize. If any-
thing, the new trends could put more pressure on earnings inequality. We
are already seeing some of this play out—for example, when we go shop-
ping and take our groceries to a kiosk instead of a cashier, or when we call a
customer service help line and interact with an automated customer service
representative.
It would be wrong, however, to believe that inequality is purely a function
of technology. Relative wages do depend in part on the demand for labor,
which is partially a function of technology. However, they also depend on
the supply of diff erent levels of skill—in other words, the distribution of
educational attainment (Goldin and Katz 2008)—and also on institutional
arrangements that aff ect wage setting, such as collective bargaining (Western
and Rosenfeld 2011).
Technology, in other words, is not destiny. Many countries have experi-
enced similar technological change as the United States, yet over the last
four decades the United States has seen both a greater increase in income
Fig. 12.4 Probability of automation by an occupation’s median hourly wage
Source: Executive Offi
ce of the President (2016).
If Automation in the Future Looks Like Automation in the Past 323
Fig. 12.5 Share of income earned by top 1 percent, 1975– 2015
Source: World Wealth and Income Database.
inequality and higher overall levels of inequality than other major advanced
economies, as shown in fi gure 12.5. When it comes to inequality—and, as I
will note in a moment, to the labor market more broadly—institutions and
policies can help determine whether and to what extent changes in tech-
nology shape economic outcomes.
12.3 The Long- Term Decline in the Labor Force Participation Rate
Raises Other Concerns about the Potential Impact of AI
Moreover, the experience of the US labor market over the last half century
raises questions around even this (relatively) optimistic view that we can
avoid large- scale job losses at the expense of greater inequality. The fact that
the labor force participation rate for men between the ages of twenty- fi ve
and fi fty- four has declined steadily from a high of 98 percent in the 1950s
to 89 percent in 2016 raises important doubts about the complacency about
full employment as a general state of the economy. As discussed in detail
in a report by the Council of Economic Advisers (2016), the decline in the
labor force participation rate has been concentrated among men with a high
school degree or less and has coincided with a decline in their relative wages.
This decline suggests that decreasing labor force participation among this
324 Jason Furman
group is a manifestation of reduced labor demand, resulting in both fewer
employment opportunities and lower wages for less- skilled men. Techno-
logical advances, including the increasing use of automation, may partly
account for this decline in demand for less- skilled labor, with globalization
likely contributing as well.
(I focus on prime- age men because I believe their experience over the past
six decades to be the best historical parallel for future eff ects of technological
change on participation in the workforce for both men and women. In the
second half of the twentieth century, prime- age women’s participation rose
sharply, as social and cultural changes in the decades following World War II
swamped any negative eff ects on participation due to technological change.
It is important to note, however, that prime- age women’s participation has
fallen in the last decade and a half—primarily for women with a high school
degree or less—paralleling the earlier experience of prime- age men.)
The concern is not that robots will take human jobs and render humans
unemployable. The traditional economic arguments against that are borne
out by centuries of experience. Instead, the concern is that the process of
turnover, in which workers displaced by technology fi nd new employment
as technology gives rise to new consumer demands and thus new jobs, could
lead to sustained periods of time with a large fraction of people not work-
ing. The traditional economic view is largely a statement about long- run
equilibrium, not about what happens in the short- to-medium term. The
fall in the labor force participation rate suggests that we must also think
carefully about short- run dynamics as the economy moves toward this long-
run equilibrium. In the short run, not all workers will have the training or
ability to fi nd the new jobs created by AI. Moreover, this “short run” (which
is a description of where the economy is in relation to equilibrium, not a
description of a defi nite length of time) could last for decades and, in fact,
the economy could be in a series of “short runs” for even longer.
As a result, AI has the potential—just like other innovations we have seen
in past decades—to contribute to further erosion in both the labor force
participation rate and the employment rate. This does not mean that we
will necessarily see a dramatically large share of jobs replaced by robots, but
even continuing on the past trend of a nearly 0.2 percentage- point annual
decline in the labor force participation rate for prime- age men would pose
substantial problems for millions of people and for the economy as a whole.
As in the case of inequality, however, we should not interpret this as
technological determinism. While most other advanced economies have
seen declines in prime- age male labor force participation, the decline in
the United States has been steeper than in almost every other advanced
economy, as shown in fi gure 12.6. Part of the reason may be that US labor
market institutions are less supportive of participation in the workforce than
other countries’ (CEA 2016).
There is no reason the economy cannot generate substantial levels of
If Automation in the Future Looks Like Automation in the Past 325
Fig. 12.6 Prime- age male labor force participation rates across the OECD
Source: Organisation for Economic Co- operation and Development.
employment at much higher levels of technology and productivity than we
have today. What matters, however, is how our labor market institutions cope
with these changes, help support the creation of new jobs, and successfully
match workers to them. Some of the potential policies along these lines
include expanding aggregate demand, increasing connective tissue in labor
markets, reforming taxes to encourage work, and creating more fl exibility
for workers. Other possible policy responses include expanding education
and training so more people have skills that complement and benefi t from
innovations, increasing the progressivity of the tax system to make sure
that everyone shares in the overall benefi ts of the economy, and expanding
institutional support for higher wages, including a higher minimum wage
and stronger collective bargaining and other forms of worker voice.
12.4 The Costs of Replacing the Current
Safety Net with a Universal Basic Income
Fears of mass job displacement as a result of automation and AI, among
other motivations, have led some to propose deep changes to the structure
of government assistance. One of the more common proposals has been
to replace some or all of the current social safety net with a universal basic
income (UBI): providing a regular, unconditional cash grant to every man,
woman, and child in the United States, instead of, say, Temporary Assis-
tance to Needy Families (TANF), the Supplemental Nutrition Assistance
Program (SNAP), or Medicaid.
While the exact contours of various UBI proposals diff er, the idea has
been put forward from the right by Charles Murray (2006), the left by Andy
Stern and Lee Kravitz (2016), and has been a staple of some technologists’
policy vision for the future (Rhodes, Krisiloff , and Altman 2016). The dif-
ferent proposals have diff erent motivations, including real and perceived
326 Jason Furman
defi ciencies in the current social safety net, the belief in a simpler and more
effi
cient system, and also the premise that we need to change our policies to
deal with the changes that will be unleashed by AI and automation more
broadly.
The issue is not that automation will render the vast majority of the popu-
lation unemployable. Instead, it is that workers will either lack the skills or
the ability to successfully match with the good, high- paying jobs created by
automation. While a market economy will do much of the work to match
workers with new job opportunities, it does not always do so successfully, as
we have seen in the past half century. Fostering skills, training, job- search
assistance, and other labor market institutions is a more direct approach to
addressing the employment issues raised by AI than UBI.
Even with these changes, however, new technologies can increase inequal-
ity and potentially even poverty through changes in the distribution of
wages. Nevertheless, replacing our current antipoverty programs with UBI
would in any realistic design make the distribution of income worse, not
better. Our tax and transfer system is largely targeted toward those in the
lower half of the income distribution, which means that it works to reduce
both poverty and income inequality. Replacing part or all of that system
with a universal cash grant, which would go to all Americans regardless of
income, would mean that relatively less of the system was targeted toward
those at the bottom—increasing, not decreasing, income inequality. Unless
one was willing to take in a much larger share of the economy in tax reve-
nues than at present, it would be diffi
cult both to provide a common amount
to all individuals and to make sure that amount was suffi
cient to cover the
needs of the poorest households. And for any additional investments in
the safety net that one would want to make, one must confront the same
targeting question.
Finally, some of the motivation for UBI has nothing to do with future
technological developments. Instead, some UBI proponents have put for-
ward the argument that it would be simpler, fairer, a
nd less distortionary
than the social assistance system we have today. This is not the space to go
into great detail on this, but suffi
ce it to say that today’s system is imperfect.
But at the same time, a wave of recent research has found that many of the
common criticisms of these programs—for example, that they discourage
work, or that they do little to reduce poverty—have been greatly overstated,
and a number of programs—including nutritional assistance, Medicaid,
and the Earned Income Tax Credit (EITC)—have important benefi ts for
the long- run earnings, health, and educational attainment of children who
grow up in recipient households.
This is not to say that we should not make the tax- and- transfer system
more progressive—just that we need to match our ambitions to the revenue
available and understand what is already successful in our social safety net.
If Automation in the Future Looks Like Automation in the Past 327
12.5 Conclusion
Artifi cial intelligence is a critical area of innovation in the US economy
right now. At least to date, AI has not had a large impact on the aggregate
performance of the macroeconomy or the labor market. But it will likely
become more important in the years to come—bringing substantial oppor-
tunities—and our fi rst impulse should be to embrace it fully.
We need more productivity growth, including through more AI. Most of
the innovation will be driven by the private sector, but government policies
also have an impact through basic research and establishing a regulatory
environment around privacy, cybersecurity, and competition.
At the same time, with or without AI we would have a lot to do if we want
to address high levels of inequality and the falling labor force participation
rate. To the degree that we are optimistic about AI, that should increase our
motivation to undertake these changes. But there is little basis for believing
that AI should dramatically change the overall direction or goals of our
current policies.
Exogenous technological developments do not uniquely determine the
future of growth, inequality, or employment. Public policy—including pub-
lic policies to help workers displaced by technology fi nd new and better jobs